]>
Commit | Line | Data |
---|---|---|
0793a61d | 1 | /* |
57c0c15b | 2 | * Performance events core code: |
0793a61d | 3 | * |
98144511 | 4 | * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de> |
e7e7ee2e IM |
5 | * Copyright (C) 2008-2011 Red Hat, Inc., Ingo Molnar |
6 | * Copyright (C) 2008-2011 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com> | |
d36b6910 | 7 | * Copyright © 2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com> |
7b732a75 | 8 | * |
57c0c15b | 9 | * For licensing details see kernel-base/COPYING |
0793a61d TG |
10 | */ |
11 | ||
12 | #include <linux/fs.h> | |
b9cacc7b | 13 | #include <linux/mm.h> |
0793a61d TG |
14 | #include <linux/cpu.h> |
15 | #include <linux/smp.h> | |
2e80a82a | 16 | #include <linux/idr.h> |
04289bb9 | 17 | #include <linux/file.h> |
0793a61d | 18 | #include <linux/poll.h> |
5a0e3ad6 | 19 | #include <linux/slab.h> |
76e1d904 | 20 | #include <linux/hash.h> |
12351ef8 | 21 | #include <linux/tick.h> |
0793a61d | 22 | #include <linux/sysfs.h> |
22a4f650 | 23 | #include <linux/dcache.h> |
0793a61d | 24 | #include <linux/percpu.h> |
22a4f650 | 25 | #include <linux/ptrace.h> |
c277443c | 26 | #include <linux/reboot.h> |
b9cacc7b | 27 | #include <linux/vmstat.h> |
abe43400 | 28 | #include <linux/device.h> |
6e5fdeed | 29 | #include <linux/export.h> |
906010b2 | 30 | #include <linux/vmalloc.h> |
b9cacc7b PZ |
31 | #include <linux/hardirq.h> |
32 | #include <linux/rculist.h> | |
0793a61d TG |
33 | #include <linux/uaccess.h> |
34 | #include <linux/syscalls.h> | |
35 | #include <linux/anon_inodes.h> | |
aa9c4c0f | 36 | #include <linux/kernel_stat.h> |
cdd6c482 | 37 | #include <linux/perf_event.h> |
6fb2915d | 38 | #include <linux/ftrace_event.h> |
3c502e7a | 39 | #include <linux/hw_breakpoint.h> |
c5ebcedb | 40 | #include <linux/mm_types.h> |
877c6856 | 41 | #include <linux/cgroup.h> |
0793a61d | 42 | |
76369139 FW |
43 | #include "internal.h" |
44 | ||
4e193bd4 TB |
45 | #include <asm/irq_regs.h> |
46 | ||
fe4b04fa | 47 | struct remote_function_call { |
e7e7ee2e IM |
48 | struct task_struct *p; |
49 | int (*func)(void *info); | |
50 | void *info; | |
51 | int ret; | |
fe4b04fa PZ |
52 | }; |
53 | ||
54 | static void remote_function(void *data) | |
55 | { | |
56 | struct remote_function_call *tfc = data; | |
57 | struct task_struct *p = tfc->p; | |
58 | ||
59 | if (p) { | |
60 | tfc->ret = -EAGAIN; | |
61 | if (task_cpu(p) != smp_processor_id() || !task_curr(p)) | |
62 | return; | |
63 | } | |
64 | ||
65 | tfc->ret = tfc->func(tfc->info); | |
66 | } | |
67 | ||
68 | /** | |
69 | * task_function_call - call a function on the cpu on which a task runs | |
70 | * @p: the task to evaluate | |
71 | * @func: the function to be called | |
72 | * @info: the function call argument | |
73 | * | |
74 | * Calls the function @func when the task is currently running. This might | |
75 | * be on the current CPU, which just calls the function directly | |
76 | * | |
77 | * returns: @func return value, or | |
78 | * -ESRCH - when the process isn't running | |
79 | * -EAGAIN - when the process moved away | |
80 | */ | |
81 | static int | |
82 | task_function_call(struct task_struct *p, int (*func) (void *info), void *info) | |
83 | { | |
84 | struct remote_function_call data = { | |
e7e7ee2e IM |
85 | .p = p, |
86 | .func = func, | |
87 | .info = info, | |
88 | .ret = -ESRCH, /* No such (running) process */ | |
fe4b04fa PZ |
89 | }; |
90 | ||
91 | if (task_curr(p)) | |
92 | smp_call_function_single(task_cpu(p), remote_function, &data, 1); | |
93 | ||
94 | return data.ret; | |
95 | } | |
96 | ||
97 | /** | |
98 | * cpu_function_call - call a function on the cpu | |
99 | * @func: the function to be called | |
100 | * @info: the function call argument | |
101 | * | |
102 | * Calls the function @func on the remote cpu. | |
103 | * | |
104 | * returns: @func return value or -ENXIO when the cpu is offline | |
105 | */ | |
106 | static int cpu_function_call(int cpu, int (*func) (void *info), void *info) | |
107 | { | |
108 | struct remote_function_call data = { | |
e7e7ee2e IM |
109 | .p = NULL, |
110 | .func = func, | |
111 | .info = info, | |
112 | .ret = -ENXIO, /* No such CPU */ | |
fe4b04fa PZ |
113 | }; |
114 | ||
115 | smp_call_function_single(cpu, remote_function, &data, 1); | |
116 | ||
117 | return data.ret; | |
118 | } | |
119 | ||
e5d1367f SE |
120 | #define PERF_FLAG_ALL (PERF_FLAG_FD_NO_GROUP |\ |
121 | PERF_FLAG_FD_OUTPUT |\ | |
122 | PERF_FLAG_PID_CGROUP) | |
123 | ||
bce38cd5 SE |
124 | /* |
125 | * branch priv levels that need permission checks | |
126 | */ | |
127 | #define PERF_SAMPLE_BRANCH_PERM_PLM \ | |
128 | (PERF_SAMPLE_BRANCH_KERNEL |\ | |
129 | PERF_SAMPLE_BRANCH_HV) | |
130 | ||
0b3fcf17 SE |
131 | enum event_type_t { |
132 | EVENT_FLEXIBLE = 0x1, | |
133 | EVENT_PINNED = 0x2, | |
134 | EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED, | |
135 | }; | |
136 | ||
e5d1367f SE |
137 | /* |
138 | * perf_sched_events : >0 events exist | |
139 | * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu | |
140 | */ | |
c5905afb | 141 | struct static_key_deferred perf_sched_events __read_mostly; |
e5d1367f | 142 | static DEFINE_PER_CPU(atomic_t, perf_cgroup_events); |
d010b332 | 143 | static DEFINE_PER_CPU(atomic_t, perf_branch_stack_events); |
e5d1367f | 144 | |
cdd6c482 IM |
145 | static atomic_t nr_mmap_events __read_mostly; |
146 | static atomic_t nr_comm_events __read_mostly; | |
147 | static atomic_t nr_task_events __read_mostly; | |
948b26b6 | 148 | static atomic_t nr_freq_events __read_mostly; |
9ee318a7 | 149 | |
108b02cf PZ |
150 | static LIST_HEAD(pmus); |
151 | static DEFINE_MUTEX(pmus_lock); | |
152 | static struct srcu_struct pmus_srcu; | |
153 | ||
0764771d | 154 | /* |
cdd6c482 | 155 | * perf event paranoia level: |
0fbdea19 IM |
156 | * -1 - not paranoid at all |
157 | * 0 - disallow raw tracepoint access for unpriv | |
cdd6c482 | 158 | * 1 - disallow cpu events for unpriv |
0fbdea19 | 159 | * 2 - disallow kernel profiling for unpriv |
0764771d | 160 | */ |
cdd6c482 | 161 | int sysctl_perf_event_paranoid __read_mostly = 1; |
0764771d | 162 | |
20443384 FW |
163 | /* Minimum for 512 kiB + 1 user control page */ |
164 | int sysctl_perf_event_mlock __read_mostly = 512 + (PAGE_SIZE / 1024); /* 'free' kiB per user */ | |
df58ab24 PZ |
165 | |
166 | /* | |
cdd6c482 | 167 | * max perf event sample rate |
df58ab24 | 168 | */ |
14c63f17 DH |
169 | #define DEFAULT_MAX_SAMPLE_RATE 100000 |
170 | #define DEFAULT_SAMPLE_PERIOD_NS (NSEC_PER_SEC / DEFAULT_MAX_SAMPLE_RATE) | |
171 | #define DEFAULT_CPU_TIME_MAX_PERCENT 25 | |
172 | ||
173 | int sysctl_perf_event_sample_rate __read_mostly = DEFAULT_MAX_SAMPLE_RATE; | |
174 | ||
175 | static int max_samples_per_tick __read_mostly = DIV_ROUND_UP(DEFAULT_MAX_SAMPLE_RATE, HZ); | |
176 | static int perf_sample_period_ns __read_mostly = DEFAULT_SAMPLE_PERIOD_NS; | |
177 | ||
178 | static atomic_t perf_sample_allowed_ns __read_mostly = | |
179 | ATOMIC_INIT( DEFAULT_SAMPLE_PERIOD_NS * DEFAULT_CPU_TIME_MAX_PERCENT / 100); | |
180 | ||
181 | void update_perf_cpu_limits(void) | |
182 | { | |
183 | u64 tmp = perf_sample_period_ns; | |
184 | ||
185 | tmp *= sysctl_perf_cpu_time_max_percent; | |
e5302920 | 186 | do_div(tmp, 100); |
14c63f17 DH |
187 | atomic_set(&perf_sample_allowed_ns, tmp); |
188 | } | |
163ec435 | 189 | |
9e630205 SE |
190 | static int perf_rotate_context(struct perf_cpu_context *cpuctx); |
191 | ||
163ec435 PZ |
192 | int perf_proc_update_handler(struct ctl_table *table, int write, |
193 | void __user *buffer, size_t *lenp, | |
194 | loff_t *ppos) | |
195 | { | |
196 | int ret = proc_dointvec(table, write, buffer, lenp, ppos); | |
197 | ||
198 | if (ret || !write) | |
199 | return ret; | |
200 | ||
201 | max_samples_per_tick = DIV_ROUND_UP(sysctl_perf_event_sample_rate, HZ); | |
14c63f17 DH |
202 | perf_sample_period_ns = NSEC_PER_SEC / sysctl_perf_event_sample_rate; |
203 | update_perf_cpu_limits(); | |
204 | ||
205 | return 0; | |
206 | } | |
207 | ||
208 | int sysctl_perf_cpu_time_max_percent __read_mostly = DEFAULT_CPU_TIME_MAX_PERCENT; | |
209 | ||
210 | int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write, | |
211 | void __user *buffer, size_t *lenp, | |
212 | loff_t *ppos) | |
213 | { | |
214 | int ret = proc_dointvec(table, write, buffer, lenp, ppos); | |
215 | ||
216 | if (ret || !write) | |
217 | return ret; | |
218 | ||
219 | update_perf_cpu_limits(); | |
163ec435 PZ |
220 | |
221 | return 0; | |
222 | } | |
1ccd1549 | 223 | |
14c63f17 DH |
224 | /* |
225 | * perf samples are done in some very critical code paths (NMIs). | |
226 | * If they take too much CPU time, the system can lock up and not | |
227 | * get any real work done. This will drop the sample rate when | |
228 | * we detect that events are taking too long. | |
229 | */ | |
230 | #define NR_ACCUMULATED_SAMPLES 128 | |
231 | DEFINE_PER_CPU(u64, running_sample_length); | |
232 | ||
233 | void perf_sample_event_took(u64 sample_len_ns) | |
234 | { | |
235 | u64 avg_local_sample_len; | |
e5302920 | 236 | u64 local_samples_len; |
14c63f17 DH |
237 | |
238 | if (atomic_read(&perf_sample_allowed_ns) == 0) | |
239 | return; | |
240 | ||
241 | /* decay the counter by 1 average sample */ | |
242 | local_samples_len = __get_cpu_var(running_sample_length); | |
243 | local_samples_len -= local_samples_len/NR_ACCUMULATED_SAMPLES; | |
244 | local_samples_len += sample_len_ns; | |
245 | __get_cpu_var(running_sample_length) = local_samples_len; | |
246 | ||
247 | /* | |
248 | * note: this will be biased artifically low until we have | |
249 | * seen NR_ACCUMULATED_SAMPLES. Doing it this way keeps us | |
250 | * from having to maintain a count. | |
251 | */ | |
252 | avg_local_sample_len = local_samples_len/NR_ACCUMULATED_SAMPLES; | |
253 | ||
254 | if (avg_local_sample_len <= atomic_read(&perf_sample_allowed_ns)) | |
255 | return; | |
256 | ||
257 | if (max_samples_per_tick <= 1) | |
258 | return; | |
259 | ||
260 | max_samples_per_tick = DIV_ROUND_UP(max_samples_per_tick, 2); | |
261 | sysctl_perf_event_sample_rate = max_samples_per_tick * HZ; | |
262 | perf_sample_period_ns = NSEC_PER_SEC / sysctl_perf_event_sample_rate; | |
263 | ||
264 | printk_ratelimited(KERN_WARNING | |
265 | "perf samples too long (%lld > %d), lowering " | |
266 | "kernel.perf_event_max_sample_rate to %d\n", | |
267 | avg_local_sample_len, | |
268 | atomic_read(&perf_sample_allowed_ns), | |
269 | sysctl_perf_event_sample_rate); | |
270 | ||
271 | update_perf_cpu_limits(); | |
272 | } | |
273 | ||
cdd6c482 | 274 | static atomic64_t perf_event_id; |
a96bbc16 | 275 | |
0b3fcf17 SE |
276 | static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx, |
277 | enum event_type_t event_type); | |
278 | ||
279 | static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx, | |
e5d1367f SE |
280 | enum event_type_t event_type, |
281 | struct task_struct *task); | |
282 | ||
283 | static void update_context_time(struct perf_event_context *ctx); | |
284 | static u64 perf_event_time(struct perf_event *event); | |
0b3fcf17 | 285 | |
cdd6c482 | 286 | void __weak perf_event_print_debug(void) { } |
0793a61d | 287 | |
84c79910 | 288 | extern __weak const char *perf_pmu_name(void) |
0793a61d | 289 | { |
84c79910 | 290 | return "pmu"; |
0793a61d TG |
291 | } |
292 | ||
0b3fcf17 SE |
293 | static inline u64 perf_clock(void) |
294 | { | |
295 | return local_clock(); | |
296 | } | |
297 | ||
e5d1367f SE |
298 | static inline struct perf_cpu_context * |
299 | __get_cpu_context(struct perf_event_context *ctx) | |
300 | { | |
301 | return this_cpu_ptr(ctx->pmu->pmu_cpu_context); | |
302 | } | |
303 | ||
facc4307 PZ |
304 | static void perf_ctx_lock(struct perf_cpu_context *cpuctx, |
305 | struct perf_event_context *ctx) | |
306 | { | |
307 | raw_spin_lock(&cpuctx->ctx.lock); | |
308 | if (ctx) | |
309 | raw_spin_lock(&ctx->lock); | |
310 | } | |
311 | ||
312 | static void perf_ctx_unlock(struct perf_cpu_context *cpuctx, | |
313 | struct perf_event_context *ctx) | |
314 | { | |
315 | if (ctx) | |
316 | raw_spin_unlock(&ctx->lock); | |
317 | raw_spin_unlock(&cpuctx->ctx.lock); | |
318 | } | |
319 | ||
e5d1367f SE |
320 | #ifdef CONFIG_CGROUP_PERF |
321 | ||
877c6856 LZ |
322 | /* |
323 | * perf_cgroup_info keeps track of time_enabled for a cgroup. | |
324 | * This is a per-cpu dynamically allocated data structure. | |
325 | */ | |
326 | struct perf_cgroup_info { | |
327 | u64 time; | |
328 | u64 timestamp; | |
329 | }; | |
330 | ||
331 | struct perf_cgroup { | |
332 | struct cgroup_subsys_state css; | |
86e213e1 | 333 | struct perf_cgroup_info __percpu *info; |
877c6856 LZ |
334 | }; |
335 | ||
3f7cce3c SE |
336 | /* |
337 | * Must ensure cgroup is pinned (css_get) before calling | |
338 | * this function. In other words, we cannot call this function | |
339 | * if there is no cgroup event for the current CPU context. | |
340 | */ | |
e5d1367f SE |
341 | static inline struct perf_cgroup * |
342 | perf_cgroup_from_task(struct task_struct *task) | |
343 | { | |
8af01f56 TH |
344 | return container_of(task_css(task, perf_subsys_id), |
345 | struct perf_cgroup, css); | |
e5d1367f SE |
346 | } |
347 | ||
348 | static inline bool | |
349 | perf_cgroup_match(struct perf_event *event) | |
350 | { | |
351 | struct perf_event_context *ctx = event->ctx; | |
352 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); | |
353 | ||
ef824fa1 TH |
354 | /* @event doesn't care about cgroup */ |
355 | if (!event->cgrp) | |
356 | return true; | |
357 | ||
358 | /* wants specific cgroup scope but @cpuctx isn't associated with any */ | |
359 | if (!cpuctx->cgrp) | |
360 | return false; | |
361 | ||
362 | /* | |
363 | * Cgroup scoping is recursive. An event enabled for a cgroup is | |
364 | * also enabled for all its descendant cgroups. If @cpuctx's | |
365 | * cgroup is a descendant of @event's (the test covers identity | |
366 | * case), it's a match. | |
367 | */ | |
368 | return cgroup_is_descendant(cpuctx->cgrp->css.cgroup, | |
369 | event->cgrp->css.cgroup); | |
e5d1367f SE |
370 | } |
371 | ||
9c5da09d | 372 | static inline bool perf_tryget_cgroup(struct perf_event *event) |
e5d1367f | 373 | { |
9c5da09d | 374 | return css_tryget(&event->cgrp->css); |
e5d1367f SE |
375 | } |
376 | ||
377 | static inline void perf_put_cgroup(struct perf_event *event) | |
378 | { | |
379 | css_put(&event->cgrp->css); | |
380 | } | |
381 | ||
382 | static inline void perf_detach_cgroup(struct perf_event *event) | |
383 | { | |
384 | perf_put_cgroup(event); | |
385 | event->cgrp = NULL; | |
386 | } | |
387 | ||
388 | static inline int is_cgroup_event(struct perf_event *event) | |
389 | { | |
390 | return event->cgrp != NULL; | |
391 | } | |
392 | ||
393 | static inline u64 perf_cgroup_event_time(struct perf_event *event) | |
394 | { | |
395 | struct perf_cgroup_info *t; | |
396 | ||
397 | t = per_cpu_ptr(event->cgrp->info, event->cpu); | |
398 | return t->time; | |
399 | } | |
400 | ||
401 | static inline void __update_cgrp_time(struct perf_cgroup *cgrp) | |
402 | { | |
403 | struct perf_cgroup_info *info; | |
404 | u64 now; | |
405 | ||
406 | now = perf_clock(); | |
407 | ||
408 | info = this_cpu_ptr(cgrp->info); | |
409 | ||
410 | info->time += now - info->timestamp; | |
411 | info->timestamp = now; | |
412 | } | |
413 | ||
414 | static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx) | |
415 | { | |
416 | struct perf_cgroup *cgrp_out = cpuctx->cgrp; | |
417 | if (cgrp_out) | |
418 | __update_cgrp_time(cgrp_out); | |
419 | } | |
420 | ||
421 | static inline void update_cgrp_time_from_event(struct perf_event *event) | |
422 | { | |
3f7cce3c SE |
423 | struct perf_cgroup *cgrp; |
424 | ||
e5d1367f | 425 | /* |
3f7cce3c SE |
426 | * ensure we access cgroup data only when needed and |
427 | * when we know the cgroup is pinned (css_get) | |
e5d1367f | 428 | */ |
3f7cce3c | 429 | if (!is_cgroup_event(event)) |
e5d1367f SE |
430 | return; |
431 | ||
3f7cce3c SE |
432 | cgrp = perf_cgroup_from_task(current); |
433 | /* | |
434 | * Do not update time when cgroup is not active | |
435 | */ | |
436 | if (cgrp == event->cgrp) | |
437 | __update_cgrp_time(event->cgrp); | |
e5d1367f SE |
438 | } |
439 | ||
440 | static inline void | |
3f7cce3c SE |
441 | perf_cgroup_set_timestamp(struct task_struct *task, |
442 | struct perf_event_context *ctx) | |
e5d1367f SE |
443 | { |
444 | struct perf_cgroup *cgrp; | |
445 | struct perf_cgroup_info *info; | |
446 | ||
3f7cce3c SE |
447 | /* |
448 | * ctx->lock held by caller | |
449 | * ensure we do not access cgroup data | |
450 | * unless we have the cgroup pinned (css_get) | |
451 | */ | |
452 | if (!task || !ctx->nr_cgroups) | |
e5d1367f SE |
453 | return; |
454 | ||
455 | cgrp = perf_cgroup_from_task(task); | |
456 | info = this_cpu_ptr(cgrp->info); | |
3f7cce3c | 457 | info->timestamp = ctx->timestamp; |
e5d1367f SE |
458 | } |
459 | ||
460 | #define PERF_CGROUP_SWOUT 0x1 /* cgroup switch out every event */ | |
461 | #define PERF_CGROUP_SWIN 0x2 /* cgroup switch in events based on task */ | |
462 | ||
463 | /* | |
464 | * reschedule events based on the cgroup constraint of task. | |
465 | * | |
466 | * mode SWOUT : schedule out everything | |
467 | * mode SWIN : schedule in based on cgroup for next | |
468 | */ | |
469 | void perf_cgroup_switch(struct task_struct *task, int mode) | |
470 | { | |
471 | struct perf_cpu_context *cpuctx; | |
472 | struct pmu *pmu; | |
473 | unsigned long flags; | |
474 | ||
475 | /* | |
476 | * disable interrupts to avoid geting nr_cgroup | |
477 | * changes via __perf_event_disable(). Also | |
478 | * avoids preemption. | |
479 | */ | |
480 | local_irq_save(flags); | |
481 | ||
482 | /* | |
483 | * we reschedule only in the presence of cgroup | |
484 | * constrained events. | |
485 | */ | |
486 | rcu_read_lock(); | |
487 | ||
488 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
e5d1367f | 489 | cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); |
95cf59ea PZ |
490 | if (cpuctx->unique_pmu != pmu) |
491 | continue; /* ensure we process each cpuctx once */ | |
e5d1367f | 492 | |
e5d1367f SE |
493 | /* |
494 | * perf_cgroup_events says at least one | |
495 | * context on this CPU has cgroup events. | |
496 | * | |
497 | * ctx->nr_cgroups reports the number of cgroup | |
498 | * events for a context. | |
499 | */ | |
500 | if (cpuctx->ctx.nr_cgroups > 0) { | |
facc4307 PZ |
501 | perf_ctx_lock(cpuctx, cpuctx->task_ctx); |
502 | perf_pmu_disable(cpuctx->ctx.pmu); | |
e5d1367f SE |
503 | |
504 | if (mode & PERF_CGROUP_SWOUT) { | |
505 | cpu_ctx_sched_out(cpuctx, EVENT_ALL); | |
506 | /* | |
507 | * must not be done before ctxswout due | |
508 | * to event_filter_match() in event_sched_out() | |
509 | */ | |
510 | cpuctx->cgrp = NULL; | |
511 | } | |
512 | ||
513 | if (mode & PERF_CGROUP_SWIN) { | |
e566b76e | 514 | WARN_ON_ONCE(cpuctx->cgrp); |
95cf59ea PZ |
515 | /* |
516 | * set cgrp before ctxsw in to allow | |
517 | * event_filter_match() to not have to pass | |
518 | * task around | |
e5d1367f SE |
519 | */ |
520 | cpuctx->cgrp = perf_cgroup_from_task(task); | |
521 | cpu_ctx_sched_in(cpuctx, EVENT_ALL, task); | |
522 | } | |
facc4307 PZ |
523 | perf_pmu_enable(cpuctx->ctx.pmu); |
524 | perf_ctx_unlock(cpuctx, cpuctx->task_ctx); | |
e5d1367f | 525 | } |
e5d1367f SE |
526 | } |
527 | ||
528 | rcu_read_unlock(); | |
529 | ||
530 | local_irq_restore(flags); | |
531 | } | |
532 | ||
a8d757ef SE |
533 | static inline void perf_cgroup_sched_out(struct task_struct *task, |
534 | struct task_struct *next) | |
e5d1367f | 535 | { |
a8d757ef SE |
536 | struct perf_cgroup *cgrp1; |
537 | struct perf_cgroup *cgrp2 = NULL; | |
538 | ||
539 | /* | |
540 | * we come here when we know perf_cgroup_events > 0 | |
541 | */ | |
542 | cgrp1 = perf_cgroup_from_task(task); | |
543 | ||
544 | /* | |
545 | * next is NULL when called from perf_event_enable_on_exec() | |
546 | * that will systematically cause a cgroup_switch() | |
547 | */ | |
548 | if (next) | |
549 | cgrp2 = perf_cgroup_from_task(next); | |
550 | ||
551 | /* | |
552 | * only schedule out current cgroup events if we know | |
553 | * that we are switching to a different cgroup. Otherwise, | |
554 | * do no touch the cgroup events. | |
555 | */ | |
556 | if (cgrp1 != cgrp2) | |
557 | perf_cgroup_switch(task, PERF_CGROUP_SWOUT); | |
e5d1367f SE |
558 | } |
559 | ||
a8d757ef SE |
560 | static inline void perf_cgroup_sched_in(struct task_struct *prev, |
561 | struct task_struct *task) | |
e5d1367f | 562 | { |
a8d757ef SE |
563 | struct perf_cgroup *cgrp1; |
564 | struct perf_cgroup *cgrp2 = NULL; | |
565 | ||
566 | /* | |
567 | * we come here when we know perf_cgroup_events > 0 | |
568 | */ | |
569 | cgrp1 = perf_cgroup_from_task(task); | |
570 | ||
571 | /* prev can never be NULL */ | |
572 | cgrp2 = perf_cgroup_from_task(prev); | |
573 | ||
574 | /* | |
575 | * only need to schedule in cgroup events if we are changing | |
576 | * cgroup during ctxsw. Cgroup events were not scheduled | |
577 | * out of ctxsw out if that was not the case. | |
578 | */ | |
579 | if (cgrp1 != cgrp2) | |
580 | perf_cgroup_switch(task, PERF_CGROUP_SWIN); | |
e5d1367f SE |
581 | } |
582 | ||
583 | static inline int perf_cgroup_connect(int fd, struct perf_event *event, | |
584 | struct perf_event_attr *attr, | |
585 | struct perf_event *group_leader) | |
586 | { | |
587 | struct perf_cgroup *cgrp; | |
588 | struct cgroup_subsys_state *css; | |
2903ff01 AV |
589 | struct fd f = fdget(fd); |
590 | int ret = 0; | |
e5d1367f | 591 | |
2903ff01 | 592 | if (!f.file) |
e5d1367f SE |
593 | return -EBADF; |
594 | ||
b77d7b60 TH |
595 | rcu_read_lock(); |
596 | ||
35cf0836 | 597 | css = css_from_dir(f.file->f_dentry, &perf_subsys); |
3db272c0 LZ |
598 | if (IS_ERR(css)) { |
599 | ret = PTR_ERR(css); | |
600 | goto out; | |
601 | } | |
e5d1367f SE |
602 | |
603 | cgrp = container_of(css, struct perf_cgroup, css); | |
604 | event->cgrp = cgrp; | |
605 | ||
f75e18cb | 606 | /* must be done before we fput() the file */ |
9c5da09d SQ |
607 | if (!perf_tryget_cgroup(event)) { |
608 | event->cgrp = NULL; | |
609 | ret = -ENOENT; | |
610 | goto out; | |
611 | } | |
f75e18cb | 612 | |
e5d1367f SE |
613 | /* |
614 | * all events in a group must monitor | |
615 | * the same cgroup because a task belongs | |
616 | * to only one perf cgroup at a time | |
617 | */ | |
618 | if (group_leader && group_leader->cgrp != cgrp) { | |
619 | perf_detach_cgroup(event); | |
620 | ret = -EINVAL; | |
e5d1367f | 621 | } |
3db272c0 | 622 | out: |
b77d7b60 | 623 | rcu_read_unlock(); |
2903ff01 | 624 | fdput(f); |
e5d1367f SE |
625 | return ret; |
626 | } | |
627 | ||
628 | static inline void | |
629 | perf_cgroup_set_shadow_time(struct perf_event *event, u64 now) | |
630 | { | |
631 | struct perf_cgroup_info *t; | |
632 | t = per_cpu_ptr(event->cgrp->info, event->cpu); | |
633 | event->shadow_ctx_time = now - t->timestamp; | |
634 | } | |
635 | ||
636 | static inline void | |
637 | perf_cgroup_defer_enabled(struct perf_event *event) | |
638 | { | |
639 | /* | |
640 | * when the current task's perf cgroup does not match | |
641 | * the event's, we need to remember to call the | |
642 | * perf_mark_enable() function the first time a task with | |
643 | * a matching perf cgroup is scheduled in. | |
644 | */ | |
645 | if (is_cgroup_event(event) && !perf_cgroup_match(event)) | |
646 | event->cgrp_defer_enabled = 1; | |
647 | } | |
648 | ||
649 | static inline void | |
650 | perf_cgroup_mark_enabled(struct perf_event *event, | |
651 | struct perf_event_context *ctx) | |
652 | { | |
653 | struct perf_event *sub; | |
654 | u64 tstamp = perf_event_time(event); | |
655 | ||
656 | if (!event->cgrp_defer_enabled) | |
657 | return; | |
658 | ||
659 | event->cgrp_defer_enabled = 0; | |
660 | ||
661 | event->tstamp_enabled = tstamp - event->total_time_enabled; | |
662 | list_for_each_entry(sub, &event->sibling_list, group_entry) { | |
663 | if (sub->state >= PERF_EVENT_STATE_INACTIVE) { | |
664 | sub->tstamp_enabled = tstamp - sub->total_time_enabled; | |
665 | sub->cgrp_defer_enabled = 0; | |
666 | } | |
667 | } | |
668 | } | |
669 | #else /* !CONFIG_CGROUP_PERF */ | |
670 | ||
671 | static inline bool | |
672 | perf_cgroup_match(struct perf_event *event) | |
673 | { | |
674 | return true; | |
675 | } | |
676 | ||
677 | static inline void perf_detach_cgroup(struct perf_event *event) | |
678 | {} | |
679 | ||
680 | static inline int is_cgroup_event(struct perf_event *event) | |
681 | { | |
682 | return 0; | |
683 | } | |
684 | ||
685 | static inline u64 perf_cgroup_event_cgrp_time(struct perf_event *event) | |
686 | { | |
687 | return 0; | |
688 | } | |
689 | ||
690 | static inline void update_cgrp_time_from_event(struct perf_event *event) | |
691 | { | |
692 | } | |
693 | ||
694 | static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx) | |
695 | { | |
696 | } | |
697 | ||
a8d757ef SE |
698 | static inline void perf_cgroup_sched_out(struct task_struct *task, |
699 | struct task_struct *next) | |
e5d1367f SE |
700 | { |
701 | } | |
702 | ||
a8d757ef SE |
703 | static inline void perf_cgroup_sched_in(struct task_struct *prev, |
704 | struct task_struct *task) | |
e5d1367f SE |
705 | { |
706 | } | |
707 | ||
708 | static inline int perf_cgroup_connect(pid_t pid, struct perf_event *event, | |
709 | struct perf_event_attr *attr, | |
710 | struct perf_event *group_leader) | |
711 | { | |
712 | return -EINVAL; | |
713 | } | |
714 | ||
715 | static inline void | |
3f7cce3c SE |
716 | perf_cgroup_set_timestamp(struct task_struct *task, |
717 | struct perf_event_context *ctx) | |
e5d1367f SE |
718 | { |
719 | } | |
720 | ||
721 | void | |
722 | perf_cgroup_switch(struct task_struct *task, struct task_struct *next) | |
723 | { | |
724 | } | |
725 | ||
726 | static inline void | |
727 | perf_cgroup_set_shadow_time(struct perf_event *event, u64 now) | |
728 | { | |
729 | } | |
730 | ||
731 | static inline u64 perf_cgroup_event_time(struct perf_event *event) | |
732 | { | |
733 | return 0; | |
734 | } | |
735 | ||
736 | static inline void | |
737 | perf_cgroup_defer_enabled(struct perf_event *event) | |
738 | { | |
739 | } | |
740 | ||
741 | static inline void | |
742 | perf_cgroup_mark_enabled(struct perf_event *event, | |
743 | struct perf_event_context *ctx) | |
744 | { | |
745 | } | |
746 | #endif | |
747 | ||
9e630205 SE |
748 | /* |
749 | * set default to be dependent on timer tick just | |
750 | * like original code | |
751 | */ | |
752 | #define PERF_CPU_HRTIMER (1000 / HZ) | |
753 | /* | |
754 | * function must be called with interrupts disbled | |
755 | */ | |
756 | static enum hrtimer_restart perf_cpu_hrtimer_handler(struct hrtimer *hr) | |
757 | { | |
758 | struct perf_cpu_context *cpuctx; | |
759 | enum hrtimer_restart ret = HRTIMER_NORESTART; | |
760 | int rotations = 0; | |
761 | ||
762 | WARN_ON(!irqs_disabled()); | |
763 | ||
764 | cpuctx = container_of(hr, struct perf_cpu_context, hrtimer); | |
765 | ||
766 | rotations = perf_rotate_context(cpuctx); | |
767 | ||
768 | /* | |
769 | * arm timer if needed | |
770 | */ | |
771 | if (rotations) { | |
772 | hrtimer_forward_now(hr, cpuctx->hrtimer_interval); | |
773 | ret = HRTIMER_RESTART; | |
774 | } | |
775 | ||
776 | return ret; | |
777 | } | |
778 | ||
779 | /* CPU is going down */ | |
780 | void perf_cpu_hrtimer_cancel(int cpu) | |
781 | { | |
782 | struct perf_cpu_context *cpuctx; | |
783 | struct pmu *pmu; | |
784 | unsigned long flags; | |
785 | ||
786 | if (WARN_ON(cpu != smp_processor_id())) | |
787 | return; | |
788 | ||
789 | local_irq_save(flags); | |
790 | ||
791 | rcu_read_lock(); | |
792 | ||
793 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
794 | cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); | |
795 | ||
796 | if (pmu->task_ctx_nr == perf_sw_context) | |
797 | continue; | |
798 | ||
799 | hrtimer_cancel(&cpuctx->hrtimer); | |
800 | } | |
801 | ||
802 | rcu_read_unlock(); | |
803 | ||
804 | local_irq_restore(flags); | |
805 | } | |
806 | ||
807 | static void __perf_cpu_hrtimer_init(struct perf_cpu_context *cpuctx, int cpu) | |
808 | { | |
809 | struct hrtimer *hr = &cpuctx->hrtimer; | |
810 | struct pmu *pmu = cpuctx->ctx.pmu; | |
62b85639 | 811 | int timer; |
9e630205 SE |
812 | |
813 | /* no multiplexing needed for SW PMU */ | |
814 | if (pmu->task_ctx_nr == perf_sw_context) | |
815 | return; | |
816 | ||
62b85639 SE |
817 | /* |
818 | * check default is sane, if not set then force to | |
819 | * default interval (1/tick) | |
820 | */ | |
821 | timer = pmu->hrtimer_interval_ms; | |
822 | if (timer < 1) | |
823 | timer = pmu->hrtimer_interval_ms = PERF_CPU_HRTIMER; | |
824 | ||
825 | cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * timer); | |
9e630205 SE |
826 | |
827 | hrtimer_init(hr, CLOCK_MONOTONIC, HRTIMER_MODE_REL_PINNED); | |
828 | hr->function = perf_cpu_hrtimer_handler; | |
829 | } | |
830 | ||
831 | static void perf_cpu_hrtimer_restart(struct perf_cpu_context *cpuctx) | |
832 | { | |
833 | struct hrtimer *hr = &cpuctx->hrtimer; | |
834 | struct pmu *pmu = cpuctx->ctx.pmu; | |
835 | ||
836 | /* not for SW PMU */ | |
837 | if (pmu->task_ctx_nr == perf_sw_context) | |
838 | return; | |
839 | ||
840 | if (hrtimer_active(hr)) | |
841 | return; | |
842 | ||
843 | if (!hrtimer_callback_running(hr)) | |
844 | __hrtimer_start_range_ns(hr, cpuctx->hrtimer_interval, | |
845 | 0, HRTIMER_MODE_REL_PINNED, 0); | |
846 | } | |
847 | ||
33696fc0 | 848 | void perf_pmu_disable(struct pmu *pmu) |
9e35ad38 | 849 | { |
33696fc0 PZ |
850 | int *count = this_cpu_ptr(pmu->pmu_disable_count); |
851 | if (!(*count)++) | |
852 | pmu->pmu_disable(pmu); | |
9e35ad38 | 853 | } |
9e35ad38 | 854 | |
33696fc0 | 855 | void perf_pmu_enable(struct pmu *pmu) |
9e35ad38 | 856 | { |
33696fc0 PZ |
857 | int *count = this_cpu_ptr(pmu->pmu_disable_count); |
858 | if (!--(*count)) | |
859 | pmu->pmu_enable(pmu); | |
9e35ad38 | 860 | } |
9e35ad38 | 861 | |
e9d2b064 PZ |
862 | static DEFINE_PER_CPU(struct list_head, rotation_list); |
863 | ||
864 | /* | |
865 | * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized | |
866 | * because they're strictly cpu affine and rotate_start is called with IRQs | |
867 | * disabled, while rotate_context is called from IRQ context. | |
868 | */ | |
108b02cf | 869 | static void perf_pmu_rotate_start(struct pmu *pmu) |
9e35ad38 | 870 | { |
108b02cf | 871 | struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); |
e9d2b064 | 872 | struct list_head *head = &__get_cpu_var(rotation_list); |
b5ab4cd5 | 873 | |
e9d2b064 | 874 | WARN_ON(!irqs_disabled()); |
b5ab4cd5 | 875 | |
d84153d6 | 876 | if (list_empty(&cpuctx->rotation_list)) |
e9d2b064 | 877 | list_add(&cpuctx->rotation_list, head); |
9e35ad38 | 878 | } |
9e35ad38 | 879 | |
cdd6c482 | 880 | static void get_ctx(struct perf_event_context *ctx) |
a63eaf34 | 881 | { |
e5289d4a | 882 | WARN_ON(!atomic_inc_not_zero(&ctx->refcount)); |
a63eaf34 PM |
883 | } |
884 | ||
cdd6c482 | 885 | static void put_ctx(struct perf_event_context *ctx) |
a63eaf34 | 886 | { |
564c2b21 PM |
887 | if (atomic_dec_and_test(&ctx->refcount)) { |
888 | if (ctx->parent_ctx) | |
889 | put_ctx(ctx->parent_ctx); | |
c93f7669 PM |
890 | if (ctx->task) |
891 | put_task_struct(ctx->task); | |
cb796ff3 | 892 | kfree_rcu(ctx, rcu_head); |
564c2b21 | 893 | } |
a63eaf34 PM |
894 | } |
895 | ||
cdd6c482 | 896 | static void unclone_ctx(struct perf_event_context *ctx) |
71a851b4 PZ |
897 | { |
898 | if (ctx->parent_ctx) { | |
899 | put_ctx(ctx->parent_ctx); | |
900 | ctx->parent_ctx = NULL; | |
901 | } | |
902 | } | |
903 | ||
6844c09d ACM |
904 | static u32 perf_event_pid(struct perf_event *event, struct task_struct *p) |
905 | { | |
906 | /* | |
907 | * only top level events have the pid namespace they were created in | |
908 | */ | |
909 | if (event->parent) | |
910 | event = event->parent; | |
911 | ||
912 | return task_tgid_nr_ns(p, event->ns); | |
913 | } | |
914 | ||
915 | static u32 perf_event_tid(struct perf_event *event, struct task_struct *p) | |
916 | { | |
917 | /* | |
918 | * only top level events have the pid namespace they were created in | |
919 | */ | |
920 | if (event->parent) | |
921 | event = event->parent; | |
922 | ||
923 | return task_pid_nr_ns(p, event->ns); | |
924 | } | |
925 | ||
7f453c24 | 926 | /* |
cdd6c482 | 927 | * If we inherit events we want to return the parent event id |
7f453c24 PZ |
928 | * to userspace. |
929 | */ | |
cdd6c482 | 930 | static u64 primary_event_id(struct perf_event *event) |
7f453c24 | 931 | { |
cdd6c482 | 932 | u64 id = event->id; |
7f453c24 | 933 | |
cdd6c482 IM |
934 | if (event->parent) |
935 | id = event->parent->id; | |
7f453c24 PZ |
936 | |
937 | return id; | |
938 | } | |
939 | ||
25346b93 | 940 | /* |
cdd6c482 | 941 | * Get the perf_event_context for a task and lock it. |
25346b93 PM |
942 | * This has to cope with with the fact that until it is locked, |
943 | * the context could get moved to another task. | |
944 | */ | |
cdd6c482 | 945 | static struct perf_event_context * |
8dc85d54 | 946 | perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags) |
25346b93 | 947 | { |
cdd6c482 | 948 | struct perf_event_context *ctx; |
25346b93 | 949 | |
9ed6060d | 950 | retry: |
058ebd0e PZ |
951 | /* |
952 | * One of the few rules of preemptible RCU is that one cannot do | |
953 | * rcu_read_unlock() while holding a scheduler (or nested) lock when | |
954 | * part of the read side critical section was preemptible -- see | |
955 | * rcu_read_unlock_special(). | |
956 | * | |
957 | * Since ctx->lock nests under rq->lock we must ensure the entire read | |
958 | * side critical section is non-preemptible. | |
959 | */ | |
960 | preempt_disable(); | |
961 | rcu_read_lock(); | |
8dc85d54 | 962 | ctx = rcu_dereference(task->perf_event_ctxp[ctxn]); |
25346b93 PM |
963 | if (ctx) { |
964 | /* | |
965 | * If this context is a clone of another, it might | |
966 | * get swapped for another underneath us by | |
cdd6c482 | 967 | * perf_event_task_sched_out, though the |
25346b93 PM |
968 | * rcu_read_lock() protects us from any context |
969 | * getting freed. Lock the context and check if it | |
970 | * got swapped before we could get the lock, and retry | |
971 | * if so. If we locked the right context, then it | |
972 | * can't get swapped on us any more. | |
973 | */ | |
e625cce1 | 974 | raw_spin_lock_irqsave(&ctx->lock, *flags); |
8dc85d54 | 975 | if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) { |
e625cce1 | 976 | raw_spin_unlock_irqrestore(&ctx->lock, *flags); |
058ebd0e PZ |
977 | rcu_read_unlock(); |
978 | preempt_enable(); | |
25346b93 PM |
979 | goto retry; |
980 | } | |
b49a9e7e PZ |
981 | |
982 | if (!atomic_inc_not_zero(&ctx->refcount)) { | |
e625cce1 | 983 | raw_spin_unlock_irqrestore(&ctx->lock, *flags); |
b49a9e7e PZ |
984 | ctx = NULL; |
985 | } | |
25346b93 PM |
986 | } |
987 | rcu_read_unlock(); | |
058ebd0e | 988 | preempt_enable(); |
25346b93 PM |
989 | return ctx; |
990 | } | |
991 | ||
992 | /* | |
993 | * Get the context for a task and increment its pin_count so it | |
994 | * can't get swapped to another task. This also increments its | |
995 | * reference count so that the context can't get freed. | |
996 | */ | |
8dc85d54 PZ |
997 | static struct perf_event_context * |
998 | perf_pin_task_context(struct task_struct *task, int ctxn) | |
25346b93 | 999 | { |
cdd6c482 | 1000 | struct perf_event_context *ctx; |
25346b93 PM |
1001 | unsigned long flags; |
1002 | ||
8dc85d54 | 1003 | ctx = perf_lock_task_context(task, ctxn, &flags); |
25346b93 PM |
1004 | if (ctx) { |
1005 | ++ctx->pin_count; | |
e625cce1 | 1006 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
25346b93 PM |
1007 | } |
1008 | return ctx; | |
1009 | } | |
1010 | ||
cdd6c482 | 1011 | static void perf_unpin_context(struct perf_event_context *ctx) |
25346b93 PM |
1012 | { |
1013 | unsigned long flags; | |
1014 | ||
e625cce1 | 1015 | raw_spin_lock_irqsave(&ctx->lock, flags); |
25346b93 | 1016 | --ctx->pin_count; |
e625cce1 | 1017 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
25346b93 PM |
1018 | } |
1019 | ||
f67218c3 PZ |
1020 | /* |
1021 | * Update the record of the current time in a context. | |
1022 | */ | |
1023 | static void update_context_time(struct perf_event_context *ctx) | |
1024 | { | |
1025 | u64 now = perf_clock(); | |
1026 | ||
1027 | ctx->time += now - ctx->timestamp; | |
1028 | ctx->timestamp = now; | |
1029 | } | |
1030 | ||
4158755d SE |
1031 | static u64 perf_event_time(struct perf_event *event) |
1032 | { | |
1033 | struct perf_event_context *ctx = event->ctx; | |
e5d1367f SE |
1034 | |
1035 | if (is_cgroup_event(event)) | |
1036 | return perf_cgroup_event_time(event); | |
1037 | ||
4158755d SE |
1038 | return ctx ? ctx->time : 0; |
1039 | } | |
1040 | ||
f67218c3 PZ |
1041 | /* |
1042 | * Update the total_time_enabled and total_time_running fields for a event. | |
b7526f0c | 1043 | * The caller of this function needs to hold the ctx->lock. |
f67218c3 PZ |
1044 | */ |
1045 | static void update_event_times(struct perf_event *event) | |
1046 | { | |
1047 | struct perf_event_context *ctx = event->ctx; | |
1048 | u64 run_end; | |
1049 | ||
1050 | if (event->state < PERF_EVENT_STATE_INACTIVE || | |
1051 | event->group_leader->state < PERF_EVENT_STATE_INACTIVE) | |
1052 | return; | |
e5d1367f SE |
1053 | /* |
1054 | * in cgroup mode, time_enabled represents | |
1055 | * the time the event was enabled AND active | |
1056 | * tasks were in the monitored cgroup. This is | |
1057 | * independent of the activity of the context as | |
1058 | * there may be a mix of cgroup and non-cgroup events. | |
1059 | * | |
1060 | * That is why we treat cgroup events differently | |
1061 | * here. | |
1062 | */ | |
1063 | if (is_cgroup_event(event)) | |
46cd6a7f | 1064 | run_end = perf_cgroup_event_time(event); |
e5d1367f SE |
1065 | else if (ctx->is_active) |
1066 | run_end = ctx->time; | |
acd1d7c1 PZ |
1067 | else |
1068 | run_end = event->tstamp_stopped; | |
1069 | ||
1070 | event->total_time_enabled = run_end - event->tstamp_enabled; | |
f67218c3 PZ |
1071 | |
1072 | if (event->state == PERF_EVENT_STATE_INACTIVE) | |
1073 | run_end = event->tstamp_stopped; | |
1074 | else | |
4158755d | 1075 | run_end = perf_event_time(event); |
f67218c3 PZ |
1076 | |
1077 | event->total_time_running = run_end - event->tstamp_running; | |
e5d1367f | 1078 | |
f67218c3 PZ |
1079 | } |
1080 | ||
96c21a46 PZ |
1081 | /* |
1082 | * Update total_time_enabled and total_time_running for all events in a group. | |
1083 | */ | |
1084 | static void update_group_times(struct perf_event *leader) | |
1085 | { | |
1086 | struct perf_event *event; | |
1087 | ||
1088 | update_event_times(leader); | |
1089 | list_for_each_entry(event, &leader->sibling_list, group_entry) | |
1090 | update_event_times(event); | |
1091 | } | |
1092 | ||
889ff015 FW |
1093 | static struct list_head * |
1094 | ctx_group_list(struct perf_event *event, struct perf_event_context *ctx) | |
1095 | { | |
1096 | if (event->attr.pinned) | |
1097 | return &ctx->pinned_groups; | |
1098 | else | |
1099 | return &ctx->flexible_groups; | |
1100 | } | |
1101 | ||
fccc714b | 1102 | /* |
cdd6c482 | 1103 | * Add a event from the lists for its context. |
fccc714b PZ |
1104 | * Must be called with ctx->mutex and ctx->lock held. |
1105 | */ | |
04289bb9 | 1106 | static void |
cdd6c482 | 1107 | list_add_event(struct perf_event *event, struct perf_event_context *ctx) |
04289bb9 | 1108 | { |
8a49542c PZ |
1109 | WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT); |
1110 | event->attach_state |= PERF_ATTACH_CONTEXT; | |
04289bb9 IM |
1111 | |
1112 | /* | |
8a49542c PZ |
1113 | * If we're a stand alone event or group leader, we go to the context |
1114 | * list, group events are kept attached to the group so that | |
1115 | * perf_group_detach can, at all times, locate all siblings. | |
04289bb9 | 1116 | */ |
8a49542c | 1117 | if (event->group_leader == event) { |
889ff015 FW |
1118 | struct list_head *list; |
1119 | ||
d6f962b5 FW |
1120 | if (is_software_event(event)) |
1121 | event->group_flags |= PERF_GROUP_SOFTWARE; | |
1122 | ||
889ff015 FW |
1123 | list = ctx_group_list(event, ctx); |
1124 | list_add_tail(&event->group_entry, list); | |
5c148194 | 1125 | } |
592903cd | 1126 | |
08309379 | 1127 | if (is_cgroup_event(event)) |
e5d1367f | 1128 | ctx->nr_cgroups++; |
e5d1367f | 1129 | |
d010b332 SE |
1130 | if (has_branch_stack(event)) |
1131 | ctx->nr_branch_stack++; | |
1132 | ||
cdd6c482 | 1133 | list_add_rcu(&event->event_entry, &ctx->event_list); |
b5ab4cd5 | 1134 | if (!ctx->nr_events) |
108b02cf | 1135 | perf_pmu_rotate_start(ctx->pmu); |
cdd6c482 IM |
1136 | ctx->nr_events++; |
1137 | if (event->attr.inherit_stat) | |
bfbd3381 | 1138 | ctx->nr_stat++; |
04289bb9 IM |
1139 | } |
1140 | ||
0231bb53 JO |
1141 | /* |
1142 | * Initialize event state based on the perf_event_attr::disabled. | |
1143 | */ | |
1144 | static inline void perf_event__state_init(struct perf_event *event) | |
1145 | { | |
1146 | event->state = event->attr.disabled ? PERF_EVENT_STATE_OFF : | |
1147 | PERF_EVENT_STATE_INACTIVE; | |
1148 | } | |
1149 | ||
c320c7b7 ACM |
1150 | /* |
1151 | * Called at perf_event creation and when events are attached/detached from a | |
1152 | * group. | |
1153 | */ | |
1154 | static void perf_event__read_size(struct perf_event *event) | |
1155 | { | |
1156 | int entry = sizeof(u64); /* value */ | |
1157 | int size = 0; | |
1158 | int nr = 1; | |
1159 | ||
1160 | if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) | |
1161 | size += sizeof(u64); | |
1162 | ||
1163 | if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
1164 | size += sizeof(u64); | |
1165 | ||
1166 | if (event->attr.read_format & PERF_FORMAT_ID) | |
1167 | entry += sizeof(u64); | |
1168 | ||
1169 | if (event->attr.read_format & PERF_FORMAT_GROUP) { | |
1170 | nr += event->group_leader->nr_siblings; | |
1171 | size += sizeof(u64); | |
1172 | } | |
1173 | ||
1174 | size += entry * nr; | |
1175 | event->read_size = size; | |
1176 | } | |
1177 | ||
1178 | static void perf_event__header_size(struct perf_event *event) | |
1179 | { | |
1180 | struct perf_sample_data *data; | |
1181 | u64 sample_type = event->attr.sample_type; | |
1182 | u16 size = 0; | |
1183 | ||
1184 | perf_event__read_size(event); | |
1185 | ||
1186 | if (sample_type & PERF_SAMPLE_IP) | |
1187 | size += sizeof(data->ip); | |
1188 | ||
6844c09d ACM |
1189 | if (sample_type & PERF_SAMPLE_ADDR) |
1190 | size += sizeof(data->addr); | |
1191 | ||
1192 | if (sample_type & PERF_SAMPLE_PERIOD) | |
1193 | size += sizeof(data->period); | |
1194 | ||
c3feedf2 AK |
1195 | if (sample_type & PERF_SAMPLE_WEIGHT) |
1196 | size += sizeof(data->weight); | |
1197 | ||
6844c09d ACM |
1198 | if (sample_type & PERF_SAMPLE_READ) |
1199 | size += event->read_size; | |
1200 | ||
d6be9ad6 SE |
1201 | if (sample_type & PERF_SAMPLE_DATA_SRC) |
1202 | size += sizeof(data->data_src.val); | |
1203 | ||
6844c09d ACM |
1204 | event->header_size = size; |
1205 | } | |
1206 | ||
1207 | static void perf_event__id_header_size(struct perf_event *event) | |
1208 | { | |
1209 | struct perf_sample_data *data; | |
1210 | u64 sample_type = event->attr.sample_type; | |
1211 | u16 size = 0; | |
1212 | ||
c320c7b7 ACM |
1213 | if (sample_type & PERF_SAMPLE_TID) |
1214 | size += sizeof(data->tid_entry); | |
1215 | ||
1216 | if (sample_type & PERF_SAMPLE_TIME) | |
1217 | size += sizeof(data->time); | |
1218 | ||
ff3d527c AH |
1219 | if (sample_type & PERF_SAMPLE_IDENTIFIER) |
1220 | size += sizeof(data->id); | |
1221 | ||
c320c7b7 ACM |
1222 | if (sample_type & PERF_SAMPLE_ID) |
1223 | size += sizeof(data->id); | |
1224 | ||
1225 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
1226 | size += sizeof(data->stream_id); | |
1227 | ||
1228 | if (sample_type & PERF_SAMPLE_CPU) | |
1229 | size += sizeof(data->cpu_entry); | |
1230 | ||
6844c09d | 1231 | event->id_header_size = size; |
c320c7b7 ACM |
1232 | } |
1233 | ||
8a49542c PZ |
1234 | static void perf_group_attach(struct perf_event *event) |
1235 | { | |
c320c7b7 | 1236 | struct perf_event *group_leader = event->group_leader, *pos; |
8a49542c | 1237 | |
74c3337c PZ |
1238 | /* |
1239 | * We can have double attach due to group movement in perf_event_open. | |
1240 | */ | |
1241 | if (event->attach_state & PERF_ATTACH_GROUP) | |
1242 | return; | |
1243 | ||
8a49542c PZ |
1244 | event->attach_state |= PERF_ATTACH_GROUP; |
1245 | ||
1246 | if (group_leader == event) | |
1247 | return; | |
1248 | ||
1249 | if (group_leader->group_flags & PERF_GROUP_SOFTWARE && | |
1250 | !is_software_event(event)) | |
1251 | group_leader->group_flags &= ~PERF_GROUP_SOFTWARE; | |
1252 | ||
1253 | list_add_tail(&event->group_entry, &group_leader->sibling_list); | |
1254 | group_leader->nr_siblings++; | |
c320c7b7 ACM |
1255 | |
1256 | perf_event__header_size(group_leader); | |
1257 | ||
1258 | list_for_each_entry(pos, &group_leader->sibling_list, group_entry) | |
1259 | perf_event__header_size(pos); | |
8a49542c PZ |
1260 | } |
1261 | ||
a63eaf34 | 1262 | /* |
cdd6c482 | 1263 | * Remove a event from the lists for its context. |
fccc714b | 1264 | * Must be called with ctx->mutex and ctx->lock held. |
a63eaf34 | 1265 | */ |
04289bb9 | 1266 | static void |
cdd6c482 | 1267 | list_del_event(struct perf_event *event, struct perf_event_context *ctx) |
04289bb9 | 1268 | { |
68cacd29 | 1269 | struct perf_cpu_context *cpuctx; |
8a49542c PZ |
1270 | /* |
1271 | * We can have double detach due to exit/hot-unplug + close. | |
1272 | */ | |
1273 | if (!(event->attach_state & PERF_ATTACH_CONTEXT)) | |
a63eaf34 | 1274 | return; |
8a49542c PZ |
1275 | |
1276 | event->attach_state &= ~PERF_ATTACH_CONTEXT; | |
1277 | ||
68cacd29 | 1278 | if (is_cgroup_event(event)) { |
e5d1367f | 1279 | ctx->nr_cgroups--; |
68cacd29 SE |
1280 | cpuctx = __get_cpu_context(ctx); |
1281 | /* | |
1282 | * if there are no more cgroup events | |
1283 | * then cler cgrp to avoid stale pointer | |
1284 | * in update_cgrp_time_from_cpuctx() | |
1285 | */ | |
1286 | if (!ctx->nr_cgroups) | |
1287 | cpuctx->cgrp = NULL; | |
1288 | } | |
e5d1367f | 1289 | |
d010b332 SE |
1290 | if (has_branch_stack(event)) |
1291 | ctx->nr_branch_stack--; | |
1292 | ||
cdd6c482 IM |
1293 | ctx->nr_events--; |
1294 | if (event->attr.inherit_stat) | |
bfbd3381 | 1295 | ctx->nr_stat--; |
8bc20959 | 1296 | |
cdd6c482 | 1297 | list_del_rcu(&event->event_entry); |
04289bb9 | 1298 | |
8a49542c PZ |
1299 | if (event->group_leader == event) |
1300 | list_del_init(&event->group_entry); | |
5c148194 | 1301 | |
96c21a46 | 1302 | update_group_times(event); |
b2e74a26 SE |
1303 | |
1304 | /* | |
1305 | * If event was in error state, then keep it | |
1306 | * that way, otherwise bogus counts will be | |
1307 | * returned on read(). The only way to get out | |
1308 | * of error state is by explicit re-enabling | |
1309 | * of the event | |
1310 | */ | |
1311 | if (event->state > PERF_EVENT_STATE_OFF) | |
1312 | event->state = PERF_EVENT_STATE_OFF; | |
050735b0 PZ |
1313 | } |
1314 | ||
8a49542c | 1315 | static void perf_group_detach(struct perf_event *event) |
050735b0 PZ |
1316 | { |
1317 | struct perf_event *sibling, *tmp; | |
8a49542c PZ |
1318 | struct list_head *list = NULL; |
1319 | ||
1320 | /* | |
1321 | * We can have double detach due to exit/hot-unplug + close. | |
1322 | */ | |
1323 | if (!(event->attach_state & PERF_ATTACH_GROUP)) | |
1324 | return; | |
1325 | ||
1326 | event->attach_state &= ~PERF_ATTACH_GROUP; | |
1327 | ||
1328 | /* | |
1329 | * If this is a sibling, remove it from its group. | |
1330 | */ | |
1331 | if (event->group_leader != event) { | |
1332 | list_del_init(&event->group_entry); | |
1333 | event->group_leader->nr_siblings--; | |
c320c7b7 | 1334 | goto out; |
8a49542c PZ |
1335 | } |
1336 | ||
1337 | if (!list_empty(&event->group_entry)) | |
1338 | list = &event->group_entry; | |
2e2af50b | 1339 | |
04289bb9 | 1340 | /* |
cdd6c482 IM |
1341 | * If this was a group event with sibling events then |
1342 | * upgrade the siblings to singleton events by adding them | |
8a49542c | 1343 | * to whatever list we are on. |
04289bb9 | 1344 | */ |
cdd6c482 | 1345 | list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) { |
8a49542c PZ |
1346 | if (list) |
1347 | list_move_tail(&sibling->group_entry, list); | |
04289bb9 | 1348 | sibling->group_leader = sibling; |
d6f962b5 FW |
1349 | |
1350 | /* Inherit group flags from the previous leader */ | |
1351 | sibling->group_flags = event->group_flags; | |
04289bb9 | 1352 | } |
c320c7b7 ACM |
1353 | |
1354 | out: | |
1355 | perf_event__header_size(event->group_leader); | |
1356 | ||
1357 | list_for_each_entry(tmp, &event->group_leader->sibling_list, group_entry) | |
1358 | perf_event__header_size(tmp); | |
04289bb9 IM |
1359 | } |
1360 | ||
fa66f07a SE |
1361 | static inline int |
1362 | event_filter_match(struct perf_event *event) | |
1363 | { | |
e5d1367f SE |
1364 | return (event->cpu == -1 || event->cpu == smp_processor_id()) |
1365 | && perf_cgroup_match(event); | |
fa66f07a SE |
1366 | } |
1367 | ||
9ffcfa6f SE |
1368 | static void |
1369 | event_sched_out(struct perf_event *event, | |
3b6f9e5c | 1370 | struct perf_cpu_context *cpuctx, |
cdd6c482 | 1371 | struct perf_event_context *ctx) |
3b6f9e5c | 1372 | { |
4158755d | 1373 | u64 tstamp = perf_event_time(event); |
fa66f07a SE |
1374 | u64 delta; |
1375 | /* | |
1376 | * An event which could not be activated because of | |
1377 | * filter mismatch still needs to have its timings | |
1378 | * maintained, otherwise bogus information is return | |
1379 | * via read() for time_enabled, time_running: | |
1380 | */ | |
1381 | if (event->state == PERF_EVENT_STATE_INACTIVE | |
1382 | && !event_filter_match(event)) { | |
e5d1367f | 1383 | delta = tstamp - event->tstamp_stopped; |
fa66f07a | 1384 | event->tstamp_running += delta; |
4158755d | 1385 | event->tstamp_stopped = tstamp; |
fa66f07a SE |
1386 | } |
1387 | ||
cdd6c482 | 1388 | if (event->state != PERF_EVENT_STATE_ACTIVE) |
9ffcfa6f | 1389 | return; |
3b6f9e5c | 1390 | |
cdd6c482 IM |
1391 | event->state = PERF_EVENT_STATE_INACTIVE; |
1392 | if (event->pending_disable) { | |
1393 | event->pending_disable = 0; | |
1394 | event->state = PERF_EVENT_STATE_OFF; | |
970892a9 | 1395 | } |
4158755d | 1396 | event->tstamp_stopped = tstamp; |
a4eaf7f1 | 1397 | event->pmu->del(event, 0); |
cdd6c482 | 1398 | event->oncpu = -1; |
3b6f9e5c | 1399 | |
cdd6c482 | 1400 | if (!is_software_event(event)) |
3b6f9e5c PM |
1401 | cpuctx->active_oncpu--; |
1402 | ctx->nr_active--; | |
0f5a2601 PZ |
1403 | if (event->attr.freq && event->attr.sample_freq) |
1404 | ctx->nr_freq--; | |
cdd6c482 | 1405 | if (event->attr.exclusive || !cpuctx->active_oncpu) |
3b6f9e5c PM |
1406 | cpuctx->exclusive = 0; |
1407 | } | |
1408 | ||
d859e29f | 1409 | static void |
cdd6c482 | 1410 | group_sched_out(struct perf_event *group_event, |
d859e29f | 1411 | struct perf_cpu_context *cpuctx, |
cdd6c482 | 1412 | struct perf_event_context *ctx) |
d859e29f | 1413 | { |
cdd6c482 | 1414 | struct perf_event *event; |
fa66f07a | 1415 | int state = group_event->state; |
d859e29f | 1416 | |
cdd6c482 | 1417 | event_sched_out(group_event, cpuctx, ctx); |
d859e29f PM |
1418 | |
1419 | /* | |
1420 | * Schedule out siblings (if any): | |
1421 | */ | |
cdd6c482 IM |
1422 | list_for_each_entry(event, &group_event->sibling_list, group_entry) |
1423 | event_sched_out(event, cpuctx, ctx); | |
d859e29f | 1424 | |
fa66f07a | 1425 | if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive) |
d859e29f PM |
1426 | cpuctx->exclusive = 0; |
1427 | } | |
1428 | ||
0793a61d | 1429 | /* |
cdd6c482 | 1430 | * Cross CPU call to remove a performance event |
0793a61d | 1431 | * |
cdd6c482 | 1432 | * We disable the event on the hardware level first. After that we |
0793a61d TG |
1433 | * remove it from the context list. |
1434 | */ | |
fe4b04fa | 1435 | static int __perf_remove_from_context(void *info) |
0793a61d | 1436 | { |
cdd6c482 IM |
1437 | struct perf_event *event = info; |
1438 | struct perf_event_context *ctx = event->ctx; | |
108b02cf | 1439 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
0793a61d | 1440 | |
e625cce1 | 1441 | raw_spin_lock(&ctx->lock); |
cdd6c482 | 1442 | event_sched_out(event, cpuctx, ctx); |
cdd6c482 | 1443 | list_del_event(event, ctx); |
64ce3126 PZ |
1444 | if (!ctx->nr_events && cpuctx->task_ctx == ctx) { |
1445 | ctx->is_active = 0; | |
1446 | cpuctx->task_ctx = NULL; | |
1447 | } | |
e625cce1 | 1448 | raw_spin_unlock(&ctx->lock); |
fe4b04fa PZ |
1449 | |
1450 | return 0; | |
0793a61d TG |
1451 | } |
1452 | ||
1453 | ||
1454 | /* | |
cdd6c482 | 1455 | * Remove the event from a task's (or a CPU's) list of events. |
0793a61d | 1456 | * |
cdd6c482 | 1457 | * CPU events are removed with a smp call. For task events we only |
0793a61d | 1458 | * call when the task is on a CPU. |
c93f7669 | 1459 | * |
cdd6c482 IM |
1460 | * If event->ctx is a cloned context, callers must make sure that |
1461 | * every task struct that event->ctx->task could possibly point to | |
c93f7669 PM |
1462 | * remains valid. This is OK when called from perf_release since |
1463 | * that only calls us on the top-level context, which can't be a clone. | |
cdd6c482 | 1464 | * When called from perf_event_exit_task, it's OK because the |
c93f7669 | 1465 | * context has been detached from its task. |
0793a61d | 1466 | */ |
fe4b04fa | 1467 | static void perf_remove_from_context(struct perf_event *event) |
0793a61d | 1468 | { |
cdd6c482 | 1469 | struct perf_event_context *ctx = event->ctx; |
0793a61d TG |
1470 | struct task_struct *task = ctx->task; |
1471 | ||
fe4b04fa PZ |
1472 | lockdep_assert_held(&ctx->mutex); |
1473 | ||
0793a61d TG |
1474 | if (!task) { |
1475 | /* | |
cdd6c482 | 1476 | * Per cpu events are removed via an smp call and |
af901ca1 | 1477 | * the removal is always successful. |
0793a61d | 1478 | */ |
fe4b04fa | 1479 | cpu_function_call(event->cpu, __perf_remove_from_context, event); |
0793a61d TG |
1480 | return; |
1481 | } | |
1482 | ||
1483 | retry: | |
fe4b04fa PZ |
1484 | if (!task_function_call(task, __perf_remove_from_context, event)) |
1485 | return; | |
0793a61d | 1486 | |
e625cce1 | 1487 | raw_spin_lock_irq(&ctx->lock); |
0793a61d | 1488 | /* |
fe4b04fa PZ |
1489 | * If we failed to find a running task, but find the context active now |
1490 | * that we've acquired the ctx->lock, retry. | |
0793a61d | 1491 | */ |
fe4b04fa | 1492 | if (ctx->is_active) { |
e625cce1 | 1493 | raw_spin_unlock_irq(&ctx->lock); |
0793a61d TG |
1494 | goto retry; |
1495 | } | |
1496 | ||
1497 | /* | |
fe4b04fa PZ |
1498 | * Since the task isn't running, its safe to remove the event, us |
1499 | * holding the ctx->lock ensures the task won't get scheduled in. | |
0793a61d | 1500 | */ |
fe4b04fa | 1501 | list_del_event(event, ctx); |
e625cce1 | 1502 | raw_spin_unlock_irq(&ctx->lock); |
0793a61d TG |
1503 | } |
1504 | ||
d859e29f | 1505 | /* |
cdd6c482 | 1506 | * Cross CPU call to disable a performance event |
d859e29f | 1507 | */ |
500ad2d8 | 1508 | int __perf_event_disable(void *info) |
d859e29f | 1509 | { |
cdd6c482 | 1510 | struct perf_event *event = info; |
cdd6c482 | 1511 | struct perf_event_context *ctx = event->ctx; |
108b02cf | 1512 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
d859e29f PM |
1513 | |
1514 | /* | |
cdd6c482 IM |
1515 | * If this is a per-task event, need to check whether this |
1516 | * event's task is the current task on this cpu. | |
fe4b04fa PZ |
1517 | * |
1518 | * Can trigger due to concurrent perf_event_context_sched_out() | |
1519 | * flipping contexts around. | |
d859e29f | 1520 | */ |
665c2142 | 1521 | if (ctx->task && cpuctx->task_ctx != ctx) |
fe4b04fa | 1522 | return -EINVAL; |
d859e29f | 1523 | |
e625cce1 | 1524 | raw_spin_lock(&ctx->lock); |
d859e29f PM |
1525 | |
1526 | /* | |
cdd6c482 | 1527 | * If the event is on, turn it off. |
d859e29f PM |
1528 | * If it is in error state, leave it in error state. |
1529 | */ | |
cdd6c482 | 1530 | if (event->state >= PERF_EVENT_STATE_INACTIVE) { |
4af4998b | 1531 | update_context_time(ctx); |
e5d1367f | 1532 | update_cgrp_time_from_event(event); |
cdd6c482 IM |
1533 | update_group_times(event); |
1534 | if (event == event->group_leader) | |
1535 | group_sched_out(event, cpuctx, ctx); | |
d859e29f | 1536 | else |
cdd6c482 IM |
1537 | event_sched_out(event, cpuctx, ctx); |
1538 | event->state = PERF_EVENT_STATE_OFF; | |
d859e29f PM |
1539 | } |
1540 | ||
e625cce1 | 1541 | raw_spin_unlock(&ctx->lock); |
fe4b04fa PZ |
1542 | |
1543 | return 0; | |
d859e29f PM |
1544 | } |
1545 | ||
1546 | /* | |
cdd6c482 | 1547 | * Disable a event. |
c93f7669 | 1548 | * |
cdd6c482 IM |
1549 | * If event->ctx is a cloned context, callers must make sure that |
1550 | * every task struct that event->ctx->task could possibly point to | |
c93f7669 | 1551 | * remains valid. This condition is satisifed when called through |
cdd6c482 IM |
1552 | * perf_event_for_each_child or perf_event_for_each because they |
1553 | * hold the top-level event's child_mutex, so any descendant that | |
1554 | * goes to exit will block in sync_child_event. | |
1555 | * When called from perf_pending_event it's OK because event->ctx | |
c93f7669 | 1556 | * is the current context on this CPU and preemption is disabled, |
cdd6c482 | 1557 | * hence we can't get into perf_event_task_sched_out for this context. |
d859e29f | 1558 | */ |
44234adc | 1559 | void perf_event_disable(struct perf_event *event) |
d859e29f | 1560 | { |
cdd6c482 | 1561 | struct perf_event_context *ctx = event->ctx; |
d859e29f PM |
1562 | struct task_struct *task = ctx->task; |
1563 | ||
1564 | if (!task) { | |
1565 | /* | |
cdd6c482 | 1566 | * Disable the event on the cpu that it's on |
d859e29f | 1567 | */ |
fe4b04fa | 1568 | cpu_function_call(event->cpu, __perf_event_disable, event); |
d859e29f PM |
1569 | return; |
1570 | } | |
1571 | ||
9ed6060d | 1572 | retry: |
fe4b04fa PZ |
1573 | if (!task_function_call(task, __perf_event_disable, event)) |
1574 | return; | |
d859e29f | 1575 | |
e625cce1 | 1576 | raw_spin_lock_irq(&ctx->lock); |
d859e29f | 1577 | /* |
cdd6c482 | 1578 | * If the event is still active, we need to retry the cross-call. |
d859e29f | 1579 | */ |
cdd6c482 | 1580 | if (event->state == PERF_EVENT_STATE_ACTIVE) { |
e625cce1 | 1581 | raw_spin_unlock_irq(&ctx->lock); |
fe4b04fa PZ |
1582 | /* |
1583 | * Reload the task pointer, it might have been changed by | |
1584 | * a concurrent perf_event_context_sched_out(). | |
1585 | */ | |
1586 | task = ctx->task; | |
d859e29f PM |
1587 | goto retry; |
1588 | } | |
1589 | ||
1590 | /* | |
1591 | * Since we have the lock this context can't be scheduled | |
1592 | * in, so we can change the state safely. | |
1593 | */ | |
cdd6c482 IM |
1594 | if (event->state == PERF_EVENT_STATE_INACTIVE) { |
1595 | update_group_times(event); | |
1596 | event->state = PERF_EVENT_STATE_OFF; | |
53cfbf59 | 1597 | } |
e625cce1 | 1598 | raw_spin_unlock_irq(&ctx->lock); |
d859e29f | 1599 | } |
dcfce4a0 | 1600 | EXPORT_SYMBOL_GPL(perf_event_disable); |
d859e29f | 1601 | |
e5d1367f SE |
1602 | static void perf_set_shadow_time(struct perf_event *event, |
1603 | struct perf_event_context *ctx, | |
1604 | u64 tstamp) | |
1605 | { | |
1606 | /* | |
1607 | * use the correct time source for the time snapshot | |
1608 | * | |
1609 | * We could get by without this by leveraging the | |
1610 | * fact that to get to this function, the caller | |
1611 | * has most likely already called update_context_time() | |
1612 | * and update_cgrp_time_xx() and thus both timestamp | |
1613 | * are identical (or very close). Given that tstamp is, | |
1614 | * already adjusted for cgroup, we could say that: | |
1615 | * tstamp - ctx->timestamp | |
1616 | * is equivalent to | |
1617 | * tstamp - cgrp->timestamp. | |
1618 | * | |
1619 | * Then, in perf_output_read(), the calculation would | |
1620 | * work with no changes because: | |
1621 | * - event is guaranteed scheduled in | |
1622 | * - no scheduled out in between | |
1623 | * - thus the timestamp would be the same | |
1624 | * | |
1625 | * But this is a bit hairy. | |
1626 | * | |
1627 | * So instead, we have an explicit cgroup call to remain | |
1628 | * within the time time source all along. We believe it | |
1629 | * is cleaner and simpler to understand. | |
1630 | */ | |
1631 | if (is_cgroup_event(event)) | |
1632 | perf_cgroup_set_shadow_time(event, tstamp); | |
1633 | else | |
1634 | event->shadow_ctx_time = tstamp - ctx->timestamp; | |
1635 | } | |
1636 | ||
4fe757dd PZ |
1637 | #define MAX_INTERRUPTS (~0ULL) |
1638 | ||
1639 | static void perf_log_throttle(struct perf_event *event, int enable); | |
1640 | ||
235c7fc7 | 1641 | static int |
9ffcfa6f | 1642 | event_sched_in(struct perf_event *event, |
235c7fc7 | 1643 | struct perf_cpu_context *cpuctx, |
6e37738a | 1644 | struct perf_event_context *ctx) |
235c7fc7 | 1645 | { |
4158755d SE |
1646 | u64 tstamp = perf_event_time(event); |
1647 | ||
cdd6c482 | 1648 | if (event->state <= PERF_EVENT_STATE_OFF) |
235c7fc7 IM |
1649 | return 0; |
1650 | ||
cdd6c482 | 1651 | event->state = PERF_EVENT_STATE_ACTIVE; |
6e37738a | 1652 | event->oncpu = smp_processor_id(); |
4fe757dd PZ |
1653 | |
1654 | /* | |
1655 | * Unthrottle events, since we scheduled we might have missed several | |
1656 | * ticks already, also for a heavily scheduling task there is little | |
1657 | * guarantee it'll get a tick in a timely manner. | |
1658 | */ | |
1659 | if (unlikely(event->hw.interrupts == MAX_INTERRUPTS)) { | |
1660 | perf_log_throttle(event, 1); | |
1661 | event->hw.interrupts = 0; | |
1662 | } | |
1663 | ||
235c7fc7 IM |
1664 | /* |
1665 | * The new state must be visible before we turn it on in the hardware: | |
1666 | */ | |
1667 | smp_wmb(); | |
1668 | ||
a4eaf7f1 | 1669 | if (event->pmu->add(event, PERF_EF_START)) { |
cdd6c482 IM |
1670 | event->state = PERF_EVENT_STATE_INACTIVE; |
1671 | event->oncpu = -1; | |
235c7fc7 IM |
1672 | return -EAGAIN; |
1673 | } | |
1674 | ||
4158755d | 1675 | event->tstamp_running += tstamp - event->tstamp_stopped; |
9ffcfa6f | 1676 | |
e5d1367f | 1677 | perf_set_shadow_time(event, ctx, tstamp); |
eed01528 | 1678 | |
cdd6c482 | 1679 | if (!is_software_event(event)) |
3b6f9e5c | 1680 | cpuctx->active_oncpu++; |
235c7fc7 | 1681 | ctx->nr_active++; |
0f5a2601 PZ |
1682 | if (event->attr.freq && event->attr.sample_freq) |
1683 | ctx->nr_freq++; | |
235c7fc7 | 1684 | |
cdd6c482 | 1685 | if (event->attr.exclusive) |
3b6f9e5c PM |
1686 | cpuctx->exclusive = 1; |
1687 | ||
235c7fc7 IM |
1688 | return 0; |
1689 | } | |
1690 | ||
6751b71e | 1691 | static int |
cdd6c482 | 1692 | group_sched_in(struct perf_event *group_event, |
6751b71e | 1693 | struct perf_cpu_context *cpuctx, |
6e37738a | 1694 | struct perf_event_context *ctx) |
6751b71e | 1695 | { |
6bde9b6c | 1696 | struct perf_event *event, *partial_group = NULL; |
51b0fe39 | 1697 | struct pmu *pmu = group_event->pmu; |
d7842da4 SE |
1698 | u64 now = ctx->time; |
1699 | bool simulate = false; | |
6751b71e | 1700 | |
cdd6c482 | 1701 | if (group_event->state == PERF_EVENT_STATE_OFF) |
6751b71e PM |
1702 | return 0; |
1703 | ||
ad5133b7 | 1704 | pmu->start_txn(pmu); |
6bde9b6c | 1705 | |
9ffcfa6f | 1706 | if (event_sched_in(group_event, cpuctx, ctx)) { |
ad5133b7 | 1707 | pmu->cancel_txn(pmu); |
9e630205 | 1708 | perf_cpu_hrtimer_restart(cpuctx); |
6751b71e | 1709 | return -EAGAIN; |
90151c35 | 1710 | } |
6751b71e PM |
1711 | |
1712 | /* | |
1713 | * Schedule in siblings as one group (if any): | |
1714 | */ | |
cdd6c482 | 1715 | list_for_each_entry(event, &group_event->sibling_list, group_entry) { |
9ffcfa6f | 1716 | if (event_sched_in(event, cpuctx, ctx)) { |
cdd6c482 | 1717 | partial_group = event; |
6751b71e PM |
1718 | goto group_error; |
1719 | } | |
1720 | } | |
1721 | ||
9ffcfa6f | 1722 | if (!pmu->commit_txn(pmu)) |
6e85158c | 1723 | return 0; |
9ffcfa6f | 1724 | |
6751b71e PM |
1725 | group_error: |
1726 | /* | |
1727 | * Groups can be scheduled in as one unit only, so undo any | |
1728 | * partial group before returning: | |
d7842da4 SE |
1729 | * The events up to the failed event are scheduled out normally, |
1730 | * tstamp_stopped will be updated. | |
1731 | * | |
1732 | * The failed events and the remaining siblings need to have | |
1733 | * their timings updated as if they had gone thru event_sched_in() | |
1734 | * and event_sched_out(). This is required to get consistent timings | |
1735 | * across the group. This also takes care of the case where the group | |
1736 | * could never be scheduled by ensuring tstamp_stopped is set to mark | |
1737 | * the time the event was actually stopped, such that time delta | |
1738 | * calculation in update_event_times() is correct. | |
6751b71e | 1739 | */ |
cdd6c482 IM |
1740 | list_for_each_entry(event, &group_event->sibling_list, group_entry) { |
1741 | if (event == partial_group) | |
d7842da4 SE |
1742 | simulate = true; |
1743 | ||
1744 | if (simulate) { | |
1745 | event->tstamp_running += now - event->tstamp_stopped; | |
1746 | event->tstamp_stopped = now; | |
1747 | } else { | |
1748 | event_sched_out(event, cpuctx, ctx); | |
1749 | } | |
6751b71e | 1750 | } |
9ffcfa6f | 1751 | event_sched_out(group_event, cpuctx, ctx); |
6751b71e | 1752 | |
ad5133b7 | 1753 | pmu->cancel_txn(pmu); |
90151c35 | 1754 | |
9e630205 SE |
1755 | perf_cpu_hrtimer_restart(cpuctx); |
1756 | ||
6751b71e PM |
1757 | return -EAGAIN; |
1758 | } | |
1759 | ||
3b6f9e5c | 1760 | /* |
cdd6c482 | 1761 | * Work out whether we can put this event group on the CPU now. |
3b6f9e5c | 1762 | */ |
cdd6c482 | 1763 | static int group_can_go_on(struct perf_event *event, |
3b6f9e5c PM |
1764 | struct perf_cpu_context *cpuctx, |
1765 | int can_add_hw) | |
1766 | { | |
1767 | /* | |
cdd6c482 | 1768 | * Groups consisting entirely of software events can always go on. |
3b6f9e5c | 1769 | */ |
d6f962b5 | 1770 | if (event->group_flags & PERF_GROUP_SOFTWARE) |
3b6f9e5c PM |
1771 | return 1; |
1772 | /* | |
1773 | * If an exclusive group is already on, no other hardware | |
cdd6c482 | 1774 | * events can go on. |
3b6f9e5c PM |
1775 | */ |
1776 | if (cpuctx->exclusive) | |
1777 | return 0; | |
1778 | /* | |
1779 | * If this group is exclusive and there are already | |
cdd6c482 | 1780 | * events on the CPU, it can't go on. |
3b6f9e5c | 1781 | */ |
cdd6c482 | 1782 | if (event->attr.exclusive && cpuctx->active_oncpu) |
3b6f9e5c PM |
1783 | return 0; |
1784 | /* | |
1785 | * Otherwise, try to add it if all previous groups were able | |
1786 | * to go on. | |
1787 | */ | |
1788 | return can_add_hw; | |
1789 | } | |
1790 | ||
cdd6c482 IM |
1791 | static void add_event_to_ctx(struct perf_event *event, |
1792 | struct perf_event_context *ctx) | |
53cfbf59 | 1793 | { |
4158755d SE |
1794 | u64 tstamp = perf_event_time(event); |
1795 | ||
cdd6c482 | 1796 | list_add_event(event, ctx); |
8a49542c | 1797 | perf_group_attach(event); |
4158755d SE |
1798 | event->tstamp_enabled = tstamp; |
1799 | event->tstamp_running = tstamp; | |
1800 | event->tstamp_stopped = tstamp; | |
53cfbf59 PM |
1801 | } |
1802 | ||
2c29ef0f PZ |
1803 | static void task_ctx_sched_out(struct perf_event_context *ctx); |
1804 | static void | |
1805 | ctx_sched_in(struct perf_event_context *ctx, | |
1806 | struct perf_cpu_context *cpuctx, | |
1807 | enum event_type_t event_type, | |
1808 | struct task_struct *task); | |
fe4b04fa | 1809 | |
dce5855b PZ |
1810 | static void perf_event_sched_in(struct perf_cpu_context *cpuctx, |
1811 | struct perf_event_context *ctx, | |
1812 | struct task_struct *task) | |
1813 | { | |
1814 | cpu_ctx_sched_in(cpuctx, EVENT_PINNED, task); | |
1815 | if (ctx) | |
1816 | ctx_sched_in(ctx, cpuctx, EVENT_PINNED, task); | |
1817 | cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, task); | |
1818 | if (ctx) | |
1819 | ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE, task); | |
1820 | } | |
1821 | ||
0793a61d | 1822 | /* |
cdd6c482 | 1823 | * Cross CPU call to install and enable a performance event |
682076ae PZ |
1824 | * |
1825 | * Must be called with ctx->mutex held | |
0793a61d | 1826 | */ |
fe4b04fa | 1827 | static int __perf_install_in_context(void *info) |
0793a61d | 1828 | { |
cdd6c482 IM |
1829 | struct perf_event *event = info; |
1830 | struct perf_event_context *ctx = event->ctx; | |
108b02cf | 1831 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
2c29ef0f PZ |
1832 | struct perf_event_context *task_ctx = cpuctx->task_ctx; |
1833 | struct task_struct *task = current; | |
1834 | ||
b58f6b0d | 1835 | perf_ctx_lock(cpuctx, task_ctx); |
2c29ef0f | 1836 | perf_pmu_disable(cpuctx->ctx.pmu); |
0793a61d TG |
1837 | |
1838 | /* | |
2c29ef0f | 1839 | * If there was an active task_ctx schedule it out. |
0793a61d | 1840 | */ |
b58f6b0d | 1841 | if (task_ctx) |
2c29ef0f | 1842 | task_ctx_sched_out(task_ctx); |
b58f6b0d PZ |
1843 | |
1844 | /* | |
1845 | * If the context we're installing events in is not the | |
1846 | * active task_ctx, flip them. | |
1847 | */ | |
1848 | if (ctx->task && task_ctx != ctx) { | |
1849 | if (task_ctx) | |
1850 | raw_spin_unlock(&task_ctx->lock); | |
1851 | raw_spin_lock(&ctx->lock); | |
1852 | task_ctx = ctx; | |
1853 | } | |
1854 | ||
1855 | if (task_ctx) { | |
1856 | cpuctx->task_ctx = task_ctx; | |
2c29ef0f PZ |
1857 | task = task_ctx->task; |
1858 | } | |
b58f6b0d | 1859 | |
2c29ef0f | 1860 | cpu_ctx_sched_out(cpuctx, EVENT_ALL); |
0793a61d | 1861 | |
4af4998b | 1862 | update_context_time(ctx); |
e5d1367f SE |
1863 | /* |
1864 | * update cgrp time only if current cgrp | |
1865 | * matches event->cgrp. Must be done before | |
1866 | * calling add_event_to_ctx() | |
1867 | */ | |
1868 | update_cgrp_time_from_event(event); | |
0793a61d | 1869 | |
cdd6c482 | 1870 | add_event_to_ctx(event, ctx); |
0793a61d | 1871 | |
d859e29f | 1872 | /* |
2c29ef0f | 1873 | * Schedule everything back in |
d859e29f | 1874 | */ |
dce5855b | 1875 | perf_event_sched_in(cpuctx, task_ctx, task); |
2c29ef0f PZ |
1876 | |
1877 | perf_pmu_enable(cpuctx->ctx.pmu); | |
1878 | perf_ctx_unlock(cpuctx, task_ctx); | |
fe4b04fa PZ |
1879 | |
1880 | return 0; | |
0793a61d TG |
1881 | } |
1882 | ||
1883 | /* | |
cdd6c482 | 1884 | * Attach a performance event to a context |
0793a61d | 1885 | * |
cdd6c482 IM |
1886 | * First we add the event to the list with the hardware enable bit |
1887 | * in event->hw_config cleared. | |
0793a61d | 1888 | * |
cdd6c482 | 1889 | * If the event is attached to a task which is on a CPU we use a smp |
0793a61d TG |
1890 | * call to enable it in the task context. The task might have been |
1891 | * scheduled away, but we check this in the smp call again. | |
1892 | */ | |
1893 | static void | |
cdd6c482 IM |
1894 | perf_install_in_context(struct perf_event_context *ctx, |
1895 | struct perf_event *event, | |
0793a61d TG |
1896 | int cpu) |
1897 | { | |
1898 | struct task_struct *task = ctx->task; | |
1899 | ||
fe4b04fa PZ |
1900 | lockdep_assert_held(&ctx->mutex); |
1901 | ||
c3f00c70 | 1902 | event->ctx = ctx; |
0cda4c02 YZ |
1903 | if (event->cpu != -1) |
1904 | event->cpu = cpu; | |
c3f00c70 | 1905 | |
0793a61d TG |
1906 | if (!task) { |
1907 | /* | |
cdd6c482 | 1908 | * Per cpu events are installed via an smp call and |
af901ca1 | 1909 | * the install is always successful. |
0793a61d | 1910 | */ |
fe4b04fa | 1911 | cpu_function_call(cpu, __perf_install_in_context, event); |
0793a61d TG |
1912 | return; |
1913 | } | |
1914 | ||
0793a61d | 1915 | retry: |
fe4b04fa PZ |
1916 | if (!task_function_call(task, __perf_install_in_context, event)) |
1917 | return; | |
0793a61d | 1918 | |
e625cce1 | 1919 | raw_spin_lock_irq(&ctx->lock); |
0793a61d | 1920 | /* |
fe4b04fa PZ |
1921 | * If we failed to find a running task, but find the context active now |
1922 | * that we've acquired the ctx->lock, retry. | |
0793a61d | 1923 | */ |
fe4b04fa | 1924 | if (ctx->is_active) { |
e625cce1 | 1925 | raw_spin_unlock_irq(&ctx->lock); |
0793a61d TG |
1926 | goto retry; |
1927 | } | |
1928 | ||
1929 | /* | |
fe4b04fa PZ |
1930 | * Since the task isn't running, its safe to add the event, us holding |
1931 | * the ctx->lock ensures the task won't get scheduled in. | |
0793a61d | 1932 | */ |
fe4b04fa | 1933 | add_event_to_ctx(event, ctx); |
e625cce1 | 1934 | raw_spin_unlock_irq(&ctx->lock); |
0793a61d TG |
1935 | } |
1936 | ||
fa289bec | 1937 | /* |
cdd6c482 | 1938 | * Put a event into inactive state and update time fields. |
fa289bec PM |
1939 | * Enabling the leader of a group effectively enables all |
1940 | * the group members that aren't explicitly disabled, so we | |
1941 | * have to update their ->tstamp_enabled also. | |
1942 | * Note: this works for group members as well as group leaders | |
1943 | * since the non-leader members' sibling_lists will be empty. | |
1944 | */ | |
1d9b482e | 1945 | static void __perf_event_mark_enabled(struct perf_event *event) |
fa289bec | 1946 | { |
cdd6c482 | 1947 | struct perf_event *sub; |
4158755d | 1948 | u64 tstamp = perf_event_time(event); |
fa289bec | 1949 | |
cdd6c482 | 1950 | event->state = PERF_EVENT_STATE_INACTIVE; |
4158755d | 1951 | event->tstamp_enabled = tstamp - event->total_time_enabled; |
9ed6060d | 1952 | list_for_each_entry(sub, &event->sibling_list, group_entry) { |
4158755d SE |
1953 | if (sub->state >= PERF_EVENT_STATE_INACTIVE) |
1954 | sub->tstamp_enabled = tstamp - sub->total_time_enabled; | |
9ed6060d | 1955 | } |
fa289bec PM |
1956 | } |
1957 | ||
d859e29f | 1958 | /* |
cdd6c482 | 1959 | * Cross CPU call to enable a performance event |
d859e29f | 1960 | */ |
fe4b04fa | 1961 | static int __perf_event_enable(void *info) |
04289bb9 | 1962 | { |
cdd6c482 | 1963 | struct perf_event *event = info; |
cdd6c482 IM |
1964 | struct perf_event_context *ctx = event->ctx; |
1965 | struct perf_event *leader = event->group_leader; | |
108b02cf | 1966 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
d859e29f | 1967 | int err; |
04289bb9 | 1968 | |
06f41796 JO |
1969 | /* |
1970 | * There's a time window between 'ctx->is_active' check | |
1971 | * in perf_event_enable function and this place having: | |
1972 | * - IRQs on | |
1973 | * - ctx->lock unlocked | |
1974 | * | |
1975 | * where the task could be killed and 'ctx' deactivated | |
1976 | * by perf_event_exit_task. | |
1977 | */ | |
1978 | if (!ctx->is_active) | |
fe4b04fa | 1979 | return -EINVAL; |
3cbed429 | 1980 | |
e625cce1 | 1981 | raw_spin_lock(&ctx->lock); |
4af4998b | 1982 | update_context_time(ctx); |
d859e29f | 1983 | |
cdd6c482 | 1984 | if (event->state >= PERF_EVENT_STATE_INACTIVE) |
d859e29f | 1985 | goto unlock; |
e5d1367f SE |
1986 | |
1987 | /* | |
1988 | * set current task's cgroup time reference point | |
1989 | */ | |
3f7cce3c | 1990 | perf_cgroup_set_timestamp(current, ctx); |
e5d1367f | 1991 | |
1d9b482e | 1992 | __perf_event_mark_enabled(event); |
04289bb9 | 1993 | |
e5d1367f SE |
1994 | if (!event_filter_match(event)) { |
1995 | if (is_cgroup_event(event)) | |
1996 | perf_cgroup_defer_enabled(event); | |
f4c4176f | 1997 | goto unlock; |
e5d1367f | 1998 | } |
f4c4176f | 1999 | |
04289bb9 | 2000 | /* |
cdd6c482 | 2001 | * If the event is in a group and isn't the group leader, |
d859e29f | 2002 | * then don't put it on unless the group is on. |
04289bb9 | 2003 | */ |
cdd6c482 | 2004 | if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE) |
d859e29f | 2005 | goto unlock; |
3b6f9e5c | 2006 | |
cdd6c482 | 2007 | if (!group_can_go_on(event, cpuctx, 1)) { |
d859e29f | 2008 | err = -EEXIST; |
e758a33d | 2009 | } else { |
cdd6c482 | 2010 | if (event == leader) |
6e37738a | 2011 | err = group_sched_in(event, cpuctx, ctx); |
e758a33d | 2012 | else |
6e37738a | 2013 | err = event_sched_in(event, cpuctx, ctx); |
e758a33d | 2014 | } |
d859e29f PM |
2015 | |
2016 | if (err) { | |
2017 | /* | |
cdd6c482 | 2018 | * If this event can't go on and it's part of a |
d859e29f PM |
2019 | * group, then the whole group has to come off. |
2020 | */ | |
9e630205 | 2021 | if (leader != event) { |
d859e29f | 2022 | group_sched_out(leader, cpuctx, ctx); |
9e630205 SE |
2023 | perf_cpu_hrtimer_restart(cpuctx); |
2024 | } | |
0d48696f | 2025 | if (leader->attr.pinned) { |
53cfbf59 | 2026 | update_group_times(leader); |
cdd6c482 | 2027 | leader->state = PERF_EVENT_STATE_ERROR; |
53cfbf59 | 2028 | } |
d859e29f PM |
2029 | } |
2030 | ||
9ed6060d | 2031 | unlock: |
e625cce1 | 2032 | raw_spin_unlock(&ctx->lock); |
fe4b04fa PZ |
2033 | |
2034 | return 0; | |
d859e29f PM |
2035 | } |
2036 | ||
2037 | /* | |
cdd6c482 | 2038 | * Enable a event. |
c93f7669 | 2039 | * |
cdd6c482 IM |
2040 | * If event->ctx is a cloned context, callers must make sure that |
2041 | * every task struct that event->ctx->task could possibly point to | |
c93f7669 | 2042 | * remains valid. This condition is satisfied when called through |
cdd6c482 IM |
2043 | * perf_event_for_each_child or perf_event_for_each as described |
2044 | * for perf_event_disable. | |
d859e29f | 2045 | */ |
44234adc | 2046 | void perf_event_enable(struct perf_event *event) |
d859e29f | 2047 | { |
cdd6c482 | 2048 | struct perf_event_context *ctx = event->ctx; |
d859e29f PM |
2049 | struct task_struct *task = ctx->task; |
2050 | ||
2051 | if (!task) { | |
2052 | /* | |
cdd6c482 | 2053 | * Enable the event on the cpu that it's on |
d859e29f | 2054 | */ |
fe4b04fa | 2055 | cpu_function_call(event->cpu, __perf_event_enable, event); |
d859e29f PM |
2056 | return; |
2057 | } | |
2058 | ||
e625cce1 | 2059 | raw_spin_lock_irq(&ctx->lock); |
cdd6c482 | 2060 | if (event->state >= PERF_EVENT_STATE_INACTIVE) |
d859e29f PM |
2061 | goto out; |
2062 | ||
2063 | /* | |
cdd6c482 IM |
2064 | * If the event is in error state, clear that first. |
2065 | * That way, if we see the event in error state below, we | |
d859e29f PM |
2066 | * know that it has gone back into error state, as distinct |
2067 | * from the task having been scheduled away before the | |
2068 | * cross-call arrived. | |
2069 | */ | |
cdd6c482 IM |
2070 | if (event->state == PERF_EVENT_STATE_ERROR) |
2071 | event->state = PERF_EVENT_STATE_OFF; | |
d859e29f | 2072 | |
9ed6060d | 2073 | retry: |
fe4b04fa | 2074 | if (!ctx->is_active) { |
1d9b482e | 2075 | __perf_event_mark_enabled(event); |
fe4b04fa PZ |
2076 | goto out; |
2077 | } | |
2078 | ||
e625cce1 | 2079 | raw_spin_unlock_irq(&ctx->lock); |
fe4b04fa PZ |
2080 | |
2081 | if (!task_function_call(task, __perf_event_enable, event)) | |
2082 | return; | |
d859e29f | 2083 | |
e625cce1 | 2084 | raw_spin_lock_irq(&ctx->lock); |
d859e29f PM |
2085 | |
2086 | /* | |
cdd6c482 | 2087 | * If the context is active and the event is still off, |
d859e29f PM |
2088 | * we need to retry the cross-call. |
2089 | */ | |
fe4b04fa PZ |
2090 | if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF) { |
2091 | /* | |
2092 | * task could have been flipped by a concurrent | |
2093 | * perf_event_context_sched_out() | |
2094 | */ | |
2095 | task = ctx->task; | |
d859e29f | 2096 | goto retry; |
fe4b04fa | 2097 | } |
fa289bec | 2098 | |
9ed6060d | 2099 | out: |
e625cce1 | 2100 | raw_spin_unlock_irq(&ctx->lock); |
d859e29f | 2101 | } |
dcfce4a0 | 2102 | EXPORT_SYMBOL_GPL(perf_event_enable); |
d859e29f | 2103 | |
26ca5c11 | 2104 | int perf_event_refresh(struct perf_event *event, int refresh) |
79f14641 | 2105 | { |
2023b359 | 2106 | /* |
cdd6c482 | 2107 | * not supported on inherited events |
2023b359 | 2108 | */ |
2e939d1d | 2109 | if (event->attr.inherit || !is_sampling_event(event)) |
2023b359 PZ |
2110 | return -EINVAL; |
2111 | ||
cdd6c482 IM |
2112 | atomic_add(refresh, &event->event_limit); |
2113 | perf_event_enable(event); | |
2023b359 PZ |
2114 | |
2115 | return 0; | |
79f14641 | 2116 | } |
26ca5c11 | 2117 | EXPORT_SYMBOL_GPL(perf_event_refresh); |
79f14641 | 2118 | |
5b0311e1 FW |
2119 | static void ctx_sched_out(struct perf_event_context *ctx, |
2120 | struct perf_cpu_context *cpuctx, | |
2121 | enum event_type_t event_type) | |
235c7fc7 | 2122 | { |
cdd6c482 | 2123 | struct perf_event *event; |
db24d33e | 2124 | int is_active = ctx->is_active; |
235c7fc7 | 2125 | |
db24d33e | 2126 | ctx->is_active &= ~event_type; |
cdd6c482 | 2127 | if (likely(!ctx->nr_events)) |
facc4307 PZ |
2128 | return; |
2129 | ||
4af4998b | 2130 | update_context_time(ctx); |
e5d1367f | 2131 | update_cgrp_time_from_cpuctx(cpuctx); |
5b0311e1 | 2132 | if (!ctx->nr_active) |
facc4307 | 2133 | return; |
5b0311e1 | 2134 | |
075e0b00 | 2135 | perf_pmu_disable(ctx->pmu); |
db24d33e | 2136 | if ((is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) { |
889ff015 FW |
2137 | list_for_each_entry(event, &ctx->pinned_groups, group_entry) |
2138 | group_sched_out(event, cpuctx, ctx); | |
9ed6060d | 2139 | } |
889ff015 | 2140 | |
db24d33e | 2141 | if ((is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) { |
889ff015 | 2142 | list_for_each_entry(event, &ctx->flexible_groups, group_entry) |
8c9ed8e1 | 2143 | group_sched_out(event, cpuctx, ctx); |
9ed6060d | 2144 | } |
1b9a644f | 2145 | perf_pmu_enable(ctx->pmu); |
235c7fc7 IM |
2146 | } |
2147 | ||
564c2b21 PM |
2148 | /* |
2149 | * Test whether two contexts are equivalent, i.e. whether they | |
2150 | * have both been cloned from the same version of the same context | |
cdd6c482 IM |
2151 | * and they both have the same number of enabled events. |
2152 | * If the number of enabled events is the same, then the set | |
2153 | * of enabled events should be the same, because these are both | |
2154 | * inherited contexts, therefore we can't access individual events | |
564c2b21 | 2155 | * in them directly with an fd; we can only enable/disable all |
cdd6c482 | 2156 | * events via prctl, or enable/disable all events in a family |
564c2b21 PM |
2157 | * via ioctl, which will have the same effect on both contexts. |
2158 | */ | |
cdd6c482 IM |
2159 | static int context_equiv(struct perf_event_context *ctx1, |
2160 | struct perf_event_context *ctx2) | |
564c2b21 PM |
2161 | { |
2162 | return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx | |
ad3a37de | 2163 | && ctx1->parent_gen == ctx2->parent_gen |
25346b93 | 2164 | && !ctx1->pin_count && !ctx2->pin_count; |
564c2b21 PM |
2165 | } |
2166 | ||
cdd6c482 IM |
2167 | static void __perf_event_sync_stat(struct perf_event *event, |
2168 | struct perf_event *next_event) | |
bfbd3381 PZ |
2169 | { |
2170 | u64 value; | |
2171 | ||
cdd6c482 | 2172 | if (!event->attr.inherit_stat) |
bfbd3381 PZ |
2173 | return; |
2174 | ||
2175 | /* | |
cdd6c482 | 2176 | * Update the event value, we cannot use perf_event_read() |
bfbd3381 PZ |
2177 | * because we're in the middle of a context switch and have IRQs |
2178 | * disabled, which upsets smp_call_function_single(), however | |
cdd6c482 | 2179 | * we know the event must be on the current CPU, therefore we |
bfbd3381 PZ |
2180 | * don't need to use it. |
2181 | */ | |
cdd6c482 IM |
2182 | switch (event->state) { |
2183 | case PERF_EVENT_STATE_ACTIVE: | |
3dbebf15 PZ |
2184 | event->pmu->read(event); |
2185 | /* fall-through */ | |
bfbd3381 | 2186 | |
cdd6c482 IM |
2187 | case PERF_EVENT_STATE_INACTIVE: |
2188 | update_event_times(event); | |
bfbd3381 PZ |
2189 | break; |
2190 | ||
2191 | default: | |
2192 | break; | |
2193 | } | |
2194 | ||
2195 | /* | |
cdd6c482 | 2196 | * In order to keep per-task stats reliable we need to flip the event |
bfbd3381 PZ |
2197 | * values when we flip the contexts. |
2198 | */ | |
e7850595 PZ |
2199 | value = local64_read(&next_event->count); |
2200 | value = local64_xchg(&event->count, value); | |
2201 | local64_set(&next_event->count, value); | |
bfbd3381 | 2202 | |
cdd6c482 IM |
2203 | swap(event->total_time_enabled, next_event->total_time_enabled); |
2204 | swap(event->total_time_running, next_event->total_time_running); | |
19d2e755 | 2205 | |
bfbd3381 | 2206 | /* |
19d2e755 | 2207 | * Since we swizzled the values, update the user visible data too. |
bfbd3381 | 2208 | */ |
cdd6c482 IM |
2209 | perf_event_update_userpage(event); |
2210 | perf_event_update_userpage(next_event); | |
bfbd3381 PZ |
2211 | } |
2212 | ||
2213 | #define list_next_entry(pos, member) \ | |
2214 | list_entry(pos->member.next, typeof(*pos), member) | |
2215 | ||
cdd6c482 IM |
2216 | static void perf_event_sync_stat(struct perf_event_context *ctx, |
2217 | struct perf_event_context *next_ctx) | |
bfbd3381 | 2218 | { |
cdd6c482 | 2219 | struct perf_event *event, *next_event; |
bfbd3381 PZ |
2220 | |
2221 | if (!ctx->nr_stat) | |
2222 | return; | |
2223 | ||
02ffdbc8 PZ |
2224 | update_context_time(ctx); |
2225 | ||
cdd6c482 IM |
2226 | event = list_first_entry(&ctx->event_list, |
2227 | struct perf_event, event_entry); | |
bfbd3381 | 2228 | |
cdd6c482 IM |
2229 | next_event = list_first_entry(&next_ctx->event_list, |
2230 | struct perf_event, event_entry); | |
bfbd3381 | 2231 | |
cdd6c482 IM |
2232 | while (&event->event_entry != &ctx->event_list && |
2233 | &next_event->event_entry != &next_ctx->event_list) { | |
bfbd3381 | 2234 | |
cdd6c482 | 2235 | __perf_event_sync_stat(event, next_event); |
bfbd3381 | 2236 | |
cdd6c482 IM |
2237 | event = list_next_entry(event, event_entry); |
2238 | next_event = list_next_entry(next_event, event_entry); | |
bfbd3381 PZ |
2239 | } |
2240 | } | |
2241 | ||
fe4b04fa PZ |
2242 | static void perf_event_context_sched_out(struct task_struct *task, int ctxn, |
2243 | struct task_struct *next) | |
0793a61d | 2244 | { |
8dc85d54 | 2245 | struct perf_event_context *ctx = task->perf_event_ctxp[ctxn]; |
cdd6c482 IM |
2246 | struct perf_event_context *next_ctx; |
2247 | struct perf_event_context *parent; | |
108b02cf | 2248 | struct perf_cpu_context *cpuctx; |
c93f7669 | 2249 | int do_switch = 1; |
0793a61d | 2250 | |
108b02cf PZ |
2251 | if (likely(!ctx)) |
2252 | return; | |
10989fb2 | 2253 | |
108b02cf PZ |
2254 | cpuctx = __get_cpu_context(ctx); |
2255 | if (!cpuctx->task_ctx) | |
0793a61d TG |
2256 | return; |
2257 | ||
c93f7669 PM |
2258 | rcu_read_lock(); |
2259 | parent = rcu_dereference(ctx->parent_ctx); | |
8dc85d54 | 2260 | next_ctx = next->perf_event_ctxp[ctxn]; |
c93f7669 PM |
2261 | if (parent && next_ctx && |
2262 | rcu_dereference(next_ctx->parent_ctx) == parent) { | |
2263 | /* | |
2264 | * Looks like the two contexts are clones, so we might be | |
2265 | * able to optimize the context switch. We lock both | |
2266 | * contexts and check that they are clones under the | |
2267 | * lock (including re-checking that neither has been | |
2268 | * uncloned in the meantime). It doesn't matter which | |
2269 | * order we take the locks because no other cpu could | |
2270 | * be trying to lock both of these tasks. | |
2271 | */ | |
e625cce1 TG |
2272 | raw_spin_lock(&ctx->lock); |
2273 | raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING); | |
c93f7669 | 2274 | if (context_equiv(ctx, next_ctx)) { |
665c2142 PZ |
2275 | /* |
2276 | * XXX do we need a memory barrier of sorts | |
cdd6c482 | 2277 | * wrt to rcu_dereference() of perf_event_ctxp |
665c2142 | 2278 | */ |
8dc85d54 PZ |
2279 | task->perf_event_ctxp[ctxn] = next_ctx; |
2280 | next->perf_event_ctxp[ctxn] = ctx; | |
c93f7669 PM |
2281 | ctx->task = next; |
2282 | next_ctx->task = task; | |
2283 | do_switch = 0; | |
bfbd3381 | 2284 | |
cdd6c482 | 2285 | perf_event_sync_stat(ctx, next_ctx); |
c93f7669 | 2286 | } |
e625cce1 TG |
2287 | raw_spin_unlock(&next_ctx->lock); |
2288 | raw_spin_unlock(&ctx->lock); | |
564c2b21 | 2289 | } |
c93f7669 | 2290 | rcu_read_unlock(); |
564c2b21 | 2291 | |
c93f7669 | 2292 | if (do_switch) { |
facc4307 | 2293 | raw_spin_lock(&ctx->lock); |
5b0311e1 | 2294 | ctx_sched_out(ctx, cpuctx, EVENT_ALL); |
c93f7669 | 2295 | cpuctx->task_ctx = NULL; |
facc4307 | 2296 | raw_spin_unlock(&ctx->lock); |
c93f7669 | 2297 | } |
0793a61d TG |
2298 | } |
2299 | ||
8dc85d54 PZ |
2300 | #define for_each_task_context_nr(ctxn) \ |
2301 | for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++) | |
2302 | ||
2303 | /* | |
2304 | * Called from scheduler to remove the events of the current task, | |
2305 | * with interrupts disabled. | |
2306 | * | |
2307 | * We stop each event and update the event value in event->count. | |
2308 | * | |
2309 | * This does not protect us against NMI, but disable() | |
2310 | * sets the disabled bit in the control field of event _before_ | |
2311 | * accessing the event control register. If a NMI hits, then it will | |
2312 | * not restart the event. | |
2313 | */ | |
ab0cce56 JO |
2314 | void __perf_event_task_sched_out(struct task_struct *task, |
2315 | struct task_struct *next) | |
8dc85d54 PZ |
2316 | { |
2317 | int ctxn; | |
2318 | ||
8dc85d54 PZ |
2319 | for_each_task_context_nr(ctxn) |
2320 | perf_event_context_sched_out(task, ctxn, next); | |
e5d1367f SE |
2321 | |
2322 | /* | |
2323 | * if cgroup events exist on this CPU, then we need | |
2324 | * to check if we have to switch out PMU state. | |
2325 | * cgroup event are system-wide mode only | |
2326 | */ | |
2327 | if (atomic_read(&__get_cpu_var(perf_cgroup_events))) | |
a8d757ef | 2328 | perf_cgroup_sched_out(task, next); |
8dc85d54 PZ |
2329 | } |
2330 | ||
04dc2dbb | 2331 | static void task_ctx_sched_out(struct perf_event_context *ctx) |
a08b159f | 2332 | { |
108b02cf | 2333 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
a08b159f | 2334 | |
a63eaf34 PM |
2335 | if (!cpuctx->task_ctx) |
2336 | return; | |
012b84da IM |
2337 | |
2338 | if (WARN_ON_ONCE(ctx != cpuctx->task_ctx)) | |
2339 | return; | |
2340 | ||
04dc2dbb | 2341 | ctx_sched_out(ctx, cpuctx, EVENT_ALL); |
a08b159f PM |
2342 | cpuctx->task_ctx = NULL; |
2343 | } | |
2344 | ||
5b0311e1 FW |
2345 | /* |
2346 | * Called with IRQs disabled | |
2347 | */ | |
2348 | static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx, | |
2349 | enum event_type_t event_type) | |
2350 | { | |
2351 | ctx_sched_out(&cpuctx->ctx, cpuctx, event_type); | |
04289bb9 IM |
2352 | } |
2353 | ||
235c7fc7 | 2354 | static void |
5b0311e1 | 2355 | ctx_pinned_sched_in(struct perf_event_context *ctx, |
6e37738a | 2356 | struct perf_cpu_context *cpuctx) |
0793a61d | 2357 | { |
cdd6c482 | 2358 | struct perf_event *event; |
0793a61d | 2359 | |
889ff015 FW |
2360 | list_for_each_entry(event, &ctx->pinned_groups, group_entry) { |
2361 | if (event->state <= PERF_EVENT_STATE_OFF) | |
3b6f9e5c | 2362 | continue; |
5632ab12 | 2363 | if (!event_filter_match(event)) |
3b6f9e5c PM |
2364 | continue; |
2365 | ||
e5d1367f SE |
2366 | /* may need to reset tstamp_enabled */ |
2367 | if (is_cgroup_event(event)) | |
2368 | perf_cgroup_mark_enabled(event, ctx); | |
2369 | ||
8c9ed8e1 | 2370 | if (group_can_go_on(event, cpuctx, 1)) |
6e37738a | 2371 | group_sched_in(event, cpuctx, ctx); |
3b6f9e5c PM |
2372 | |
2373 | /* | |
2374 | * If this pinned group hasn't been scheduled, | |
2375 | * put it in error state. | |
2376 | */ | |
cdd6c482 IM |
2377 | if (event->state == PERF_EVENT_STATE_INACTIVE) { |
2378 | update_group_times(event); | |
2379 | event->state = PERF_EVENT_STATE_ERROR; | |
53cfbf59 | 2380 | } |
3b6f9e5c | 2381 | } |
5b0311e1 FW |
2382 | } |
2383 | ||
2384 | static void | |
2385 | ctx_flexible_sched_in(struct perf_event_context *ctx, | |
6e37738a | 2386 | struct perf_cpu_context *cpuctx) |
5b0311e1 FW |
2387 | { |
2388 | struct perf_event *event; | |
2389 | int can_add_hw = 1; | |
3b6f9e5c | 2390 | |
889ff015 FW |
2391 | list_for_each_entry(event, &ctx->flexible_groups, group_entry) { |
2392 | /* Ignore events in OFF or ERROR state */ | |
2393 | if (event->state <= PERF_EVENT_STATE_OFF) | |
3b6f9e5c | 2394 | continue; |
04289bb9 IM |
2395 | /* |
2396 | * Listen to the 'cpu' scheduling filter constraint | |
cdd6c482 | 2397 | * of events: |
04289bb9 | 2398 | */ |
5632ab12 | 2399 | if (!event_filter_match(event)) |
0793a61d TG |
2400 | continue; |
2401 | ||
e5d1367f SE |
2402 | /* may need to reset tstamp_enabled */ |
2403 | if (is_cgroup_event(event)) | |
2404 | perf_cgroup_mark_enabled(event, ctx); | |
2405 | ||
9ed6060d | 2406 | if (group_can_go_on(event, cpuctx, can_add_hw)) { |
6e37738a | 2407 | if (group_sched_in(event, cpuctx, ctx)) |
dd0e6ba2 | 2408 | can_add_hw = 0; |
9ed6060d | 2409 | } |
0793a61d | 2410 | } |
5b0311e1 FW |
2411 | } |
2412 | ||
2413 | static void | |
2414 | ctx_sched_in(struct perf_event_context *ctx, | |
2415 | struct perf_cpu_context *cpuctx, | |
e5d1367f SE |
2416 | enum event_type_t event_type, |
2417 | struct task_struct *task) | |
5b0311e1 | 2418 | { |
e5d1367f | 2419 | u64 now; |
db24d33e | 2420 | int is_active = ctx->is_active; |
e5d1367f | 2421 | |
db24d33e | 2422 | ctx->is_active |= event_type; |
5b0311e1 | 2423 | if (likely(!ctx->nr_events)) |
facc4307 | 2424 | return; |
5b0311e1 | 2425 | |
e5d1367f SE |
2426 | now = perf_clock(); |
2427 | ctx->timestamp = now; | |
3f7cce3c | 2428 | perf_cgroup_set_timestamp(task, ctx); |
5b0311e1 FW |
2429 | /* |
2430 | * First go through the list and put on any pinned groups | |
2431 | * in order to give them the best chance of going on. | |
2432 | */ | |
db24d33e | 2433 | if (!(is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) |
6e37738a | 2434 | ctx_pinned_sched_in(ctx, cpuctx); |
5b0311e1 FW |
2435 | |
2436 | /* Then walk through the lower prio flexible groups */ | |
db24d33e | 2437 | if (!(is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) |
6e37738a | 2438 | ctx_flexible_sched_in(ctx, cpuctx); |
235c7fc7 IM |
2439 | } |
2440 | ||
329c0e01 | 2441 | static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx, |
e5d1367f SE |
2442 | enum event_type_t event_type, |
2443 | struct task_struct *task) | |
329c0e01 FW |
2444 | { |
2445 | struct perf_event_context *ctx = &cpuctx->ctx; | |
2446 | ||
e5d1367f | 2447 | ctx_sched_in(ctx, cpuctx, event_type, task); |
329c0e01 FW |
2448 | } |
2449 | ||
e5d1367f SE |
2450 | static void perf_event_context_sched_in(struct perf_event_context *ctx, |
2451 | struct task_struct *task) | |
235c7fc7 | 2452 | { |
108b02cf | 2453 | struct perf_cpu_context *cpuctx; |
235c7fc7 | 2454 | |
108b02cf | 2455 | cpuctx = __get_cpu_context(ctx); |
329c0e01 FW |
2456 | if (cpuctx->task_ctx == ctx) |
2457 | return; | |
2458 | ||
facc4307 | 2459 | perf_ctx_lock(cpuctx, ctx); |
1b9a644f | 2460 | perf_pmu_disable(ctx->pmu); |
329c0e01 FW |
2461 | /* |
2462 | * We want to keep the following priority order: | |
2463 | * cpu pinned (that don't need to move), task pinned, | |
2464 | * cpu flexible, task flexible. | |
2465 | */ | |
2466 | cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE); | |
2467 | ||
1d5f003f GN |
2468 | if (ctx->nr_events) |
2469 | cpuctx->task_ctx = ctx; | |
9b33fa6b | 2470 | |
86b47c25 GN |
2471 | perf_event_sched_in(cpuctx, cpuctx->task_ctx, task); |
2472 | ||
facc4307 PZ |
2473 | perf_pmu_enable(ctx->pmu); |
2474 | perf_ctx_unlock(cpuctx, ctx); | |
2475 | ||
b5ab4cd5 PZ |
2476 | /* |
2477 | * Since these rotations are per-cpu, we need to ensure the | |
2478 | * cpu-context we got scheduled on is actually rotating. | |
2479 | */ | |
108b02cf | 2480 | perf_pmu_rotate_start(ctx->pmu); |
235c7fc7 IM |
2481 | } |
2482 | ||
d010b332 SE |
2483 | /* |
2484 | * When sampling the branck stack in system-wide, it may be necessary | |
2485 | * to flush the stack on context switch. This happens when the branch | |
2486 | * stack does not tag its entries with the pid of the current task. | |
2487 | * Otherwise it becomes impossible to associate a branch entry with a | |
2488 | * task. This ambiguity is more likely to appear when the branch stack | |
2489 | * supports priv level filtering and the user sets it to monitor only | |
2490 | * at the user level (which could be a useful measurement in system-wide | |
2491 | * mode). In that case, the risk is high of having a branch stack with | |
2492 | * branch from multiple tasks. Flushing may mean dropping the existing | |
2493 | * entries or stashing them somewhere in the PMU specific code layer. | |
2494 | * | |
2495 | * This function provides the context switch callback to the lower code | |
2496 | * layer. It is invoked ONLY when there is at least one system-wide context | |
2497 | * with at least one active event using taken branch sampling. | |
2498 | */ | |
2499 | static void perf_branch_stack_sched_in(struct task_struct *prev, | |
2500 | struct task_struct *task) | |
2501 | { | |
2502 | struct perf_cpu_context *cpuctx; | |
2503 | struct pmu *pmu; | |
2504 | unsigned long flags; | |
2505 | ||
2506 | /* no need to flush branch stack if not changing task */ | |
2507 | if (prev == task) | |
2508 | return; | |
2509 | ||
2510 | local_irq_save(flags); | |
2511 | ||
2512 | rcu_read_lock(); | |
2513 | ||
2514 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
2515 | cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); | |
2516 | ||
2517 | /* | |
2518 | * check if the context has at least one | |
2519 | * event using PERF_SAMPLE_BRANCH_STACK | |
2520 | */ | |
2521 | if (cpuctx->ctx.nr_branch_stack > 0 | |
2522 | && pmu->flush_branch_stack) { | |
2523 | ||
2524 | pmu = cpuctx->ctx.pmu; | |
2525 | ||
2526 | perf_ctx_lock(cpuctx, cpuctx->task_ctx); | |
2527 | ||
2528 | perf_pmu_disable(pmu); | |
2529 | ||
2530 | pmu->flush_branch_stack(); | |
2531 | ||
2532 | perf_pmu_enable(pmu); | |
2533 | ||
2534 | perf_ctx_unlock(cpuctx, cpuctx->task_ctx); | |
2535 | } | |
2536 | } | |
2537 | ||
2538 | rcu_read_unlock(); | |
2539 | ||
2540 | local_irq_restore(flags); | |
2541 | } | |
2542 | ||
8dc85d54 PZ |
2543 | /* |
2544 | * Called from scheduler to add the events of the current task | |
2545 | * with interrupts disabled. | |
2546 | * | |
2547 | * We restore the event value and then enable it. | |
2548 | * | |
2549 | * This does not protect us against NMI, but enable() | |
2550 | * sets the enabled bit in the control field of event _before_ | |
2551 | * accessing the event control register. If a NMI hits, then it will | |
2552 | * keep the event running. | |
2553 | */ | |
ab0cce56 JO |
2554 | void __perf_event_task_sched_in(struct task_struct *prev, |
2555 | struct task_struct *task) | |
8dc85d54 PZ |
2556 | { |
2557 | struct perf_event_context *ctx; | |
2558 | int ctxn; | |
2559 | ||
2560 | for_each_task_context_nr(ctxn) { | |
2561 | ctx = task->perf_event_ctxp[ctxn]; | |
2562 | if (likely(!ctx)) | |
2563 | continue; | |
2564 | ||
e5d1367f | 2565 | perf_event_context_sched_in(ctx, task); |
8dc85d54 | 2566 | } |
e5d1367f SE |
2567 | /* |
2568 | * if cgroup events exist on this CPU, then we need | |
2569 | * to check if we have to switch in PMU state. | |
2570 | * cgroup event are system-wide mode only | |
2571 | */ | |
2572 | if (atomic_read(&__get_cpu_var(perf_cgroup_events))) | |
a8d757ef | 2573 | perf_cgroup_sched_in(prev, task); |
d010b332 SE |
2574 | |
2575 | /* check for system-wide branch_stack events */ | |
2576 | if (atomic_read(&__get_cpu_var(perf_branch_stack_events))) | |
2577 | perf_branch_stack_sched_in(prev, task); | |
235c7fc7 IM |
2578 | } |
2579 | ||
abd50713 PZ |
2580 | static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count) |
2581 | { | |
2582 | u64 frequency = event->attr.sample_freq; | |
2583 | u64 sec = NSEC_PER_SEC; | |
2584 | u64 divisor, dividend; | |
2585 | ||
2586 | int count_fls, nsec_fls, frequency_fls, sec_fls; | |
2587 | ||
2588 | count_fls = fls64(count); | |
2589 | nsec_fls = fls64(nsec); | |
2590 | frequency_fls = fls64(frequency); | |
2591 | sec_fls = 30; | |
2592 | ||
2593 | /* | |
2594 | * We got @count in @nsec, with a target of sample_freq HZ | |
2595 | * the target period becomes: | |
2596 | * | |
2597 | * @count * 10^9 | |
2598 | * period = ------------------- | |
2599 | * @nsec * sample_freq | |
2600 | * | |
2601 | */ | |
2602 | ||
2603 | /* | |
2604 | * Reduce accuracy by one bit such that @a and @b converge | |
2605 | * to a similar magnitude. | |
2606 | */ | |
fe4b04fa | 2607 | #define REDUCE_FLS(a, b) \ |
abd50713 PZ |
2608 | do { \ |
2609 | if (a##_fls > b##_fls) { \ | |
2610 | a >>= 1; \ | |
2611 | a##_fls--; \ | |
2612 | } else { \ | |
2613 | b >>= 1; \ | |
2614 | b##_fls--; \ | |
2615 | } \ | |
2616 | } while (0) | |
2617 | ||
2618 | /* | |
2619 | * Reduce accuracy until either term fits in a u64, then proceed with | |
2620 | * the other, so that finally we can do a u64/u64 division. | |
2621 | */ | |
2622 | while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) { | |
2623 | REDUCE_FLS(nsec, frequency); | |
2624 | REDUCE_FLS(sec, count); | |
2625 | } | |
2626 | ||
2627 | if (count_fls + sec_fls > 64) { | |
2628 | divisor = nsec * frequency; | |
2629 | ||
2630 | while (count_fls + sec_fls > 64) { | |
2631 | REDUCE_FLS(count, sec); | |
2632 | divisor >>= 1; | |
2633 | } | |
2634 | ||
2635 | dividend = count * sec; | |
2636 | } else { | |
2637 | dividend = count * sec; | |
2638 | ||
2639 | while (nsec_fls + frequency_fls > 64) { | |
2640 | REDUCE_FLS(nsec, frequency); | |
2641 | dividend >>= 1; | |
2642 | } | |
2643 | ||
2644 | divisor = nsec * frequency; | |
2645 | } | |
2646 | ||
f6ab91ad PZ |
2647 | if (!divisor) |
2648 | return dividend; | |
2649 | ||
abd50713 PZ |
2650 | return div64_u64(dividend, divisor); |
2651 | } | |
2652 | ||
e050e3f0 SE |
2653 | static DEFINE_PER_CPU(int, perf_throttled_count); |
2654 | static DEFINE_PER_CPU(u64, perf_throttled_seq); | |
2655 | ||
f39d47ff | 2656 | static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count, bool disable) |
bd2b5b12 | 2657 | { |
cdd6c482 | 2658 | struct hw_perf_event *hwc = &event->hw; |
f6ab91ad | 2659 | s64 period, sample_period; |
bd2b5b12 PZ |
2660 | s64 delta; |
2661 | ||
abd50713 | 2662 | period = perf_calculate_period(event, nsec, count); |
bd2b5b12 PZ |
2663 | |
2664 | delta = (s64)(period - hwc->sample_period); | |
2665 | delta = (delta + 7) / 8; /* low pass filter */ | |
2666 | ||
2667 | sample_period = hwc->sample_period + delta; | |
2668 | ||
2669 | if (!sample_period) | |
2670 | sample_period = 1; | |
2671 | ||
bd2b5b12 | 2672 | hwc->sample_period = sample_period; |
abd50713 | 2673 | |
e7850595 | 2674 | if (local64_read(&hwc->period_left) > 8*sample_period) { |
f39d47ff SE |
2675 | if (disable) |
2676 | event->pmu->stop(event, PERF_EF_UPDATE); | |
2677 | ||
e7850595 | 2678 | local64_set(&hwc->period_left, 0); |
f39d47ff SE |
2679 | |
2680 | if (disable) | |
2681 | event->pmu->start(event, PERF_EF_RELOAD); | |
abd50713 | 2682 | } |
bd2b5b12 PZ |
2683 | } |
2684 | ||
e050e3f0 SE |
2685 | /* |
2686 | * combine freq adjustment with unthrottling to avoid two passes over the | |
2687 | * events. At the same time, make sure, having freq events does not change | |
2688 | * the rate of unthrottling as that would introduce bias. | |
2689 | */ | |
2690 | static void perf_adjust_freq_unthr_context(struct perf_event_context *ctx, | |
2691 | int needs_unthr) | |
60db5e09 | 2692 | { |
cdd6c482 IM |
2693 | struct perf_event *event; |
2694 | struct hw_perf_event *hwc; | |
e050e3f0 | 2695 | u64 now, period = TICK_NSEC; |
abd50713 | 2696 | s64 delta; |
60db5e09 | 2697 | |
e050e3f0 SE |
2698 | /* |
2699 | * only need to iterate over all events iff: | |
2700 | * - context have events in frequency mode (needs freq adjust) | |
2701 | * - there are events to unthrottle on this cpu | |
2702 | */ | |
2703 | if (!(ctx->nr_freq || needs_unthr)) | |
0f5a2601 PZ |
2704 | return; |
2705 | ||
e050e3f0 | 2706 | raw_spin_lock(&ctx->lock); |
f39d47ff | 2707 | perf_pmu_disable(ctx->pmu); |
e050e3f0 | 2708 | |
03541f8b | 2709 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { |
cdd6c482 | 2710 | if (event->state != PERF_EVENT_STATE_ACTIVE) |
60db5e09 PZ |
2711 | continue; |
2712 | ||
5632ab12 | 2713 | if (!event_filter_match(event)) |
5d27c23d PZ |
2714 | continue; |
2715 | ||
cdd6c482 | 2716 | hwc = &event->hw; |
6a24ed6c | 2717 | |
ae23bff1 | 2718 | if (hwc->interrupts == MAX_INTERRUPTS) { |
e050e3f0 | 2719 | hwc->interrupts = 0; |
cdd6c482 | 2720 | perf_log_throttle(event, 1); |
a4eaf7f1 | 2721 | event->pmu->start(event, 0); |
a78ac325 PZ |
2722 | } |
2723 | ||
cdd6c482 | 2724 | if (!event->attr.freq || !event->attr.sample_freq) |
60db5e09 PZ |
2725 | continue; |
2726 | ||
e050e3f0 SE |
2727 | /* |
2728 | * stop the event and update event->count | |
2729 | */ | |
2730 | event->pmu->stop(event, PERF_EF_UPDATE); | |
2731 | ||
e7850595 | 2732 | now = local64_read(&event->count); |
abd50713 PZ |
2733 | delta = now - hwc->freq_count_stamp; |
2734 | hwc->freq_count_stamp = now; | |
60db5e09 | 2735 | |
e050e3f0 SE |
2736 | /* |
2737 | * restart the event | |
2738 | * reload only if value has changed | |
f39d47ff SE |
2739 | * we have stopped the event so tell that |
2740 | * to perf_adjust_period() to avoid stopping it | |
2741 | * twice. | |
e050e3f0 | 2742 | */ |
abd50713 | 2743 | if (delta > 0) |
f39d47ff | 2744 | perf_adjust_period(event, period, delta, false); |
e050e3f0 SE |
2745 | |
2746 | event->pmu->start(event, delta > 0 ? PERF_EF_RELOAD : 0); | |
60db5e09 | 2747 | } |
e050e3f0 | 2748 | |
f39d47ff | 2749 | perf_pmu_enable(ctx->pmu); |
e050e3f0 | 2750 | raw_spin_unlock(&ctx->lock); |
60db5e09 PZ |
2751 | } |
2752 | ||
235c7fc7 | 2753 | /* |
cdd6c482 | 2754 | * Round-robin a context's events: |
235c7fc7 | 2755 | */ |
cdd6c482 | 2756 | static void rotate_ctx(struct perf_event_context *ctx) |
0793a61d | 2757 | { |
dddd3379 TG |
2758 | /* |
2759 | * Rotate the first entry last of non-pinned groups. Rotation might be | |
2760 | * disabled by the inheritance code. | |
2761 | */ | |
2762 | if (!ctx->rotate_disable) | |
2763 | list_rotate_left(&ctx->flexible_groups); | |
235c7fc7 IM |
2764 | } |
2765 | ||
b5ab4cd5 | 2766 | /* |
e9d2b064 PZ |
2767 | * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized |
2768 | * because they're strictly cpu affine and rotate_start is called with IRQs | |
2769 | * disabled, while rotate_context is called from IRQ context. | |
b5ab4cd5 | 2770 | */ |
9e630205 | 2771 | static int perf_rotate_context(struct perf_cpu_context *cpuctx) |
235c7fc7 | 2772 | { |
8dc85d54 | 2773 | struct perf_event_context *ctx = NULL; |
e050e3f0 | 2774 | int rotate = 0, remove = 1; |
7fc23a53 | 2775 | |
b5ab4cd5 | 2776 | if (cpuctx->ctx.nr_events) { |
e9d2b064 | 2777 | remove = 0; |
b5ab4cd5 PZ |
2778 | if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active) |
2779 | rotate = 1; | |
2780 | } | |
235c7fc7 | 2781 | |
8dc85d54 | 2782 | ctx = cpuctx->task_ctx; |
b5ab4cd5 | 2783 | if (ctx && ctx->nr_events) { |
e9d2b064 | 2784 | remove = 0; |
b5ab4cd5 PZ |
2785 | if (ctx->nr_events != ctx->nr_active) |
2786 | rotate = 1; | |
2787 | } | |
9717e6cd | 2788 | |
e050e3f0 | 2789 | if (!rotate) |
0f5a2601 PZ |
2790 | goto done; |
2791 | ||
facc4307 | 2792 | perf_ctx_lock(cpuctx, cpuctx->task_ctx); |
1b9a644f | 2793 | perf_pmu_disable(cpuctx->ctx.pmu); |
60db5e09 | 2794 | |
e050e3f0 SE |
2795 | cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE); |
2796 | if (ctx) | |
2797 | ctx_sched_out(ctx, cpuctx, EVENT_FLEXIBLE); | |
0793a61d | 2798 | |
e050e3f0 SE |
2799 | rotate_ctx(&cpuctx->ctx); |
2800 | if (ctx) | |
2801 | rotate_ctx(ctx); | |
235c7fc7 | 2802 | |
e050e3f0 | 2803 | perf_event_sched_in(cpuctx, ctx, current); |
235c7fc7 | 2804 | |
0f5a2601 PZ |
2805 | perf_pmu_enable(cpuctx->ctx.pmu); |
2806 | perf_ctx_unlock(cpuctx, cpuctx->task_ctx); | |
b5ab4cd5 | 2807 | done: |
e9d2b064 PZ |
2808 | if (remove) |
2809 | list_del_init(&cpuctx->rotation_list); | |
9e630205 SE |
2810 | |
2811 | return rotate; | |
e9d2b064 PZ |
2812 | } |
2813 | ||
026249ef FW |
2814 | #ifdef CONFIG_NO_HZ_FULL |
2815 | bool perf_event_can_stop_tick(void) | |
2816 | { | |
948b26b6 | 2817 | if (atomic_read(&nr_freq_events) || |
d84153d6 | 2818 | __this_cpu_read(perf_throttled_count)) |
026249ef | 2819 | return false; |
d84153d6 FW |
2820 | else |
2821 | return true; | |
026249ef FW |
2822 | } |
2823 | #endif | |
2824 | ||
e9d2b064 PZ |
2825 | void perf_event_task_tick(void) |
2826 | { | |
2827 | struct list_head *head = &__get_cpu_var(rotation_list); | |
2828 | struct perf_cpu_context *cpuctx, *tmp; | |
e050e3f0 SE |
2829 | struct perf_event_context *ctx; |
2830 | int throttled; | |
b5ab4cd5 | 2831 | |
e9d2b064 PZ |
2832 | WARN_ON(!irqs_disabled()); |
2833 | ||
e050e3f0 SE |
2834 | __this_cpu_inc(perf_throttled_seq); |
2835 | throttled = __this_cpu_xchg(perf_throttled_count, 0); | |
2836 | ||
e9d2b064 | 2837 | list_for_each_entry_safe(cpuctx, tmp, head, rotation_list) { |
e050e3f0 SE |
2838 | ctx = &cpuctx->ctx; |
2839 | perf_adjust_freq_unthr_context(ctx, throttled); | |
2840 | ||
2841 | ctx = cpuctx->task_ctx; | |
2842 | if (ctx) | |
2843 | perf_adjust_freq_unthr_context(ctx, throttled); | |
e9d2b064 | 2844 | } |
0793a61d TG |
2845 | } |
2846 | ||
889ff015 FW |
2847 | static int event_enable_on_exec(struct perf_event *event, |
2848 | struct perf_event_context *ctx) | |
2849 | { | |
2850 | if (!event->attr.enable_on_exec) | |
2851 | return 0; | |
2852 | ||
2853 | event->attr.enable_on_exec = 0; | |
2854 | if (event->state >= PERF_EVENT_STATE_INACTIVE) | |
2855 | return 0; | |
2856 | ||
1d9b482e | 2857 | __perf_event_mark_enabled(event); |
889ff015 FW |
2858 | |
2859 | return 1; | |
2860 | } | |
2861 | ||
57e7986e | 2862 | /* |
cdd6c482 | 2863 | * Enable all of a task's events that have been marked enable-on-exec. |
57e7986e PM |
2864 | * This expects task == current. |
2865 | */ | |
8dc85d54 | 2866 | static void perf_event_enable_on_exec(struct perf_event_context *ctx) |
57e7986e | 2867 | { |
cdd6c482 | 2868 | struct perf_event *event; |
57e7986e PM |
2869 | unsigned long flags; |
2870 | int enabled = 0; | |
889ff015 | 2871 | int ret; |
57e7986e PM |
2872 | |
2873 | local_irq_save(flags); | |
cdd6c482 | 2874 | if (!ctx || !ctx->nr_events) |
57e7986e PM |
2875 | goto out; |
2876 | ||
e566b76e SE |
2877 | /* |
2878 | * We must ctxsw out cgroup events to avoid conflict | |
2879 | * when invoking perf_task_event_sched_in() later on | |
2880 | * in this function. Otherwise we end up trying to | |
2881 | * ctxswin cgroup events which are already scheduled | |
2882 | * in. | |
2883 | */ | |
a8d757ef | 2884 | perf_cgroup_sched_out(current, NULL); |
57e7986e | 2885 | |
e625cce1 | 2886 | raw_spin_lock(&ctx->lock); |
04dc2dbb | 2887 | task_ctx_sched_out(ctx); |
57e7986e | 2888 | |
b79387ef | 2889 | list_for_each_entry(event, &ctx->event_list, event_entry) { |
889ff015 FW |
2890 | ret = event_enable_on_exec(event, ctx); |
2891 | if (ret) | |
2892 | enabled = 1; | |
57e7986e PM |
2893 | } |
2894 | ||
2895 | /* | |
cdd6c482 | 2896 | * Unclone this context if we enabled any event. |
57e7986e | 2897 | */ |
71a851b4 PZ |
2898 | if (enabled) |
2899 | unclone_ctx(ctx); | |
57e7986e | 2900 | |
e625cce1 | 2901 | raw_spin_unlock(&ctx->lock); |
57e7986e | 2902 | |
e566b76e SE |
2903 | /* |
2904 | * Also calls ctxswin for cgroup events, if any: | |
2905 | */ | |
e5d1367f | 2906 | perf_event_context_sched_in(ctx, ctx->task); |
9ed6060d | 2907 | out: |
57e7986e PM |
2908 | local_irq_restore(flags); |
2909 | } | |
2910 | ||
0793a61d | 2911 | /* |
cdd6c482 | 2912 | * Cross CPU call to read the hardware event |
0793a61d | 2913 | */ |
cdd6c482 | 2914 | static void __perf_event_read(void *info) |
0793a61d | 2915 | { |
cdd6c482 IM |
2916 | struct perf_event *event = info; |
2917 | struct perf_event_context *ctx = event->ctx; | |
108b02cf | 2918 | struct perf_cpu_context *cpuctx = __get_cpu_context(ctx); |
621a01ea | 2919 | |
e1ac3614 PM |
2920 | /* |
2921 | * If this is a task context, we need to check whether it is | |
2922 | * the current task context of this cpu. If not it has been | |
2923 | * scheduled out before the smp call arrived. In that case | |
cdd6c482 IM |
2924 | * event->count would have been updated to a recent sample |
2925 | * when the event was scheduled out. | |
e1ac3614 PM |
2926 | */ |
2927 | if (ctx->task && cpuctx->task_ctx != ctx) | |
2928 | return; | |
2929 | ||
e625cce1 | 2930 | raw_spin_lock(&ctx->lock); |
e5d1367f | 2931 | if (ctx->is_active) { |
542e72fc | 2932 | update_context_time(ctx); |
e5d1367f SE |
2933 | update_cgrp_time_from_event(event); |
2934 | } | |
cdd6c482 | 2935 | update_event_times(event); |
542e72fc PZ |
2936 | if (event->state == PERF_EVENT_STATE_ACTIVE) |
2937 | event->pmu->read(event); | |
e625cce1 | 2938 | raw_spin_unlock(&ctx->lock); |
0793a61d TG |
2939 | } |
2940 | ||
b5e58793 PZ |
2941 | static inline u64 perf_event_count(struct perf_event *event) |
2942 | { | |
e7850595 | 2943 | return local64_read(&event->count) + atomic64_read(&event->child_count); |
b5e58793 PZ |
2944 | } |
2945 | ||
cdd6c482 | 2946 | static u64 perf_event_read(struct perf_event *event) |
0793a61d TG |
2947 | { |
2948 | /* | |
cdd6c482 IM |
2949 | * If event is enabled and currently active on a CPU, update the |
2950 | * value in the event structure: | |
0793a61d | 2951 | */ |
cdd6c482 IM |
2952 | if (event->state == PERF_EVENT_STATE_ACTIVE) { |
2953 | smp_call_function_single(event->oncpu, | |
2954 | __perf_event_read, event, 1); | |
2955 | } else if (event->state == PERF_EVENT_STATE_INACTIVE) { | |
2b8988c9 PZ |
2956 | struct perf_event_context *ctx = event->ctx; |
2957 | unsigned long flags; | |
2958 | ||
e625cce1 | 2959 | raw_spin_lock_irqsave(&ctx->lock, flags); |
c530ccd9 SE |
2960 | /* |
2961 | * may read while context is not active | |
2962 | * (e.g., thread is blocked), in that case | |
2963 | * we cannot update context time | |
2964 | */ | |
e5d1367f | 2965 | if (ctx->is_active) { |
c530ccd9 | 2966 | update_context_time(ctx); |
e5d1367f SE |
2967 | update_cgrp_time_from_event(event); |
2968 | } | |
cdd6c482 | 2969 | update_event_times(event); |
e625cce1 | 2970 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
0793a61d TG |
2971 | } |
2972 | ||
b5e58793 | 2973 | return perf_event_count(event); |
0793a61d TG |
2974 | } |
2975 | ||
a63eaf34 | 2976 | /* |
cdd6c482 | 2977 | * Initialize the perf_event context in a task_struct: |
a63eaf34 | 2978 | */ |
eb184479 | 2979 | static void __perf_event_init_context(struct perf_event_context *ctx) |
a63eaf34 | 2980 | { |
e625cce1 | 2981 | raw_spin_lock_init(&ctx->lock); |
a63eaf34 | 2982 | mutex_init(&ctx->mutex); |
889ff015 FW |
2983 | INIT_LIST_HEAD(&ctx->pinned_groups); |
2984 | INIT_LIST_HEAD(&ctx->flexible_groups); | |
a63eaf34 PM |
2985 | INIT_LIST_HEAD(&ctx->event_list); |
2986 | atomic_set(&ctx->refcount, 1); | |
eb184479 PZ |
2987 | } |
2988 | ||
2989 | static struct perf_event_context * | |
2990 | alloc_perf_context(struct pmu *pmu, struct task_struct *task) | |
2991 | { | |
2992 | struct perf_event_context *ctx; | |
2993 | ||
2994 | ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL); | |
2995 | if (!ctx) | |
2996 | return NULL; | |
2997 | ||
2998 | __perf_event_init_context(ctx); | |
2999 | if (task) { | |
3000 | ctx->task = task; | |
3001 | get_task_struct(task); | |
0793a61d | 3002 | } |
eb184479 PZ |
3003 | ctx->pmu = pmu; |
3004 | ||
3005 | return ctx; | |
a63eaf34 PM |
3006 | } |
3007 | ||
2ebd4ffb MH |
3008 | static struct task_struct * |
3009 | find_lively_task_by_vpid(pid_t vpid) | |
3010 | { | |
3011 | struct task_struct *task; | |
3012 | int err; | |
0793a61d TG |
3013 | |
3014 | rcu_read_lock(); | |
2ebd4ffb | 3015 | if (!vpid) |
0793a61d TG |
3016 | task = current; |
3017 | else | |
2ebd4ffb | 3018 | task = find_task_by_vpid(vpid); |
0793a61d TG |
3019 | if (task) |
3020 | get_task_struct(task); | |
3021 | rcu_read_unlock(); | |
3022 | ||
3023 | if (!task) | |
3024 | return ERR_PTR(-ESRCH); | |
3025 | ||
0793a61d | 3026 | /* Reuse ptrace permission checks for now. */ |
c93f7669 PM |
3027 | err = -EACCES; |
3028 | if (!ptrace_may_access(task, PTRACE_MODE_READ)) | |
3029 | goto errout; | |
3030 | ||
2ebd4ffb MH |
3031 | return task; |
3032 | errout: | |
3033 | put_task_struct(task); | |
3034 | return ERR_PTR(err); | |
3035 | ||
3036 | } | |
3037 | ||
fe4b04fa PZ |
3038 | /* |
3039 | * Returns a matching context with refcount and pincount. | |
3040 | */ | |
108b02cf | 3041 | static struct perf_event_context * |
38a81da2 | 3042 | find_get_context(struct pmu *pmu, struct task_struct *task, int cpu) |
0793a61d | 3043 | { |
cdd6c482 | 3044 | struct perf_event_context *ctx; |
22a4f650 | 3045 | struct perf_cpu_context *cpuctx; |
25346b93 | 3046 | unsigned long flags; |
8dc85d54 | 3047 | int ctxn, err; |
0793a61d | 3048 | |
22a4ec72 | 3049 | if (!task) { |
cdd6c482 | 3050 | /* Must be root to operate on a CPU event: */ |
0764771d | 3051 | if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN)) |
0793a61d TG |
3052 | return ERR_PTR(-EACCES); |
3053 | ||
0793a61d | 3054 | /* |
cdd6c482 | 3055 | * We could be clever and allow to attach a event to an |
0793a61d TG |
3056 | * offline CPU and activate it when the CPU comes up, but |
3057 | * that's for later. | |
3058 | */ | |
f6325e30 | 3059 | if (!cpu_online(cpu)) |
0793a61d TG |
3060 | return ERR_PTR(-ENODEV); |
3061 | ||
108b02cf | 3062 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); |
0793a61d | 3063 | ctx = &cpuctx->ctx; |
c93f7669 | 3064 | get_ctx(ctx); |
fe4b04fa | 3065 | ++ctx->pin_count; |
0793a61d | 3066 | |
0793a61d TG |
3067 | return ctx; |
3068 | } | |
3069 | ||
8dc85d54 PZ |
3070 | err = -EINVAL; |
3071 | ctxn = pmu->task_ctx_nr; | |
3072 | if (ctxn < 0) | |
3073 | goto errout; | |
3074 | ||
9ed6060d | 3075 | retry: |
8dc85d54 | 3076 | ctx = perf_lock_task_context(task, ctxn, &flags); |
c93f7669 | 3077 | if (ctx) { |
71a851b4 | 3078 | unclone_ctx(ctx); |
fe4b04fa | 3079 | ++ctx->pin_count; |
e625cce1 | 3080 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
9137fb28 | 3081 | } else { |
eb184479 | 3082 | ctx = alloc_perf_context(pmu, task); |
c93f7669 PM |
3083 | err = -ENOMEM; |
3084 | if (!ctx) | |
3085 | goto errout; | |
eb184479 | 3086 | |
dbe08d82 ON |
3087 | err = 0; |
3088 | mutex_lock(&task->perf_event_mutex); | |
3089 | /* | |
3090 | * If it has already passed perf_event_exit_task(). | |
3091 | * we must see PF_EXITING, it takes this mutex too. | |
3092 | */ | |
3093 | if (task->flags & PF_EXITING) | |
3094 | err = -ESRCH; | |
3095 | else if (task->perf_event_ctxp[ctxn]) | |
3096 | err = -EAGAIN; | |
fe4b04fa | 3097 | else { |
9137fb28 | 3098 | get_ctx(ctx); |
fe4b04fa | 3099 | ++ctx->pin_count; |
dbe08d82 | 3100 | rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx); |
fe4b04fa | 3101 | } |
dbe08d82 ON |
3102 | mutex_unlock(&task->perf_event_mutex); |
3103 | ||
3104 | if (unlikely(err)) { | |
9137fb28 | 3105 | put_ctx(ctx); |
dbe08d82 ON |
3106 | |
3107 | if (err == -EAGAIN) | |
3108 | goto retry; | |
3109 | goto errout; | |
a63eaf34 PM |
3110 | } |
3111 | } | |
3112 | ||
0793a61d | 3113 | return ctx; |
c93f7669 | 3114 | |
9ed6060d | 3115 | errout: |
c93f7669 | 3116 | return ERR_PTR(err); |
0793a61d TG |
3117 | } |
3118 | ||
6fb2915d LZ |
3119 | static void perf_event_free_filter(struct perf_event *event); |
3120 | ||
cdd6c482 | 3121 | static void free_event_rcu(struct rcu_head *head) |
592903cd | 3122 | { |
cdd6c482 | 3123 | struct perf_event *event; |
592903cd | 3124 | |
cdd6c482 IM |
3125 | event = container_of(head, struct perf_event, rcu_head); |
3126 | if (event->ns) | |
3127 | put_pid_ns(event->ns); | |
6fb2915d | 3128 | perf_event_free_filter(event); |
cdd6c482 | 3129 | kfree(event); |
592903cd PZ |
3130 | } |
3131 | ||
76369139 | 3132 | static void ring_buffer_put(struct ring_buffer *rb); |
9bb5d40c | 3133 | static void ring_buffer_detach(struct perf_event *event, struct ring_buffer *rb); |
925d519a | 3134 | |
4beb31f3 | 3135 | static void unaccount_event_cpu(struct perf_event *event, int cpu) |
f1600952 | 3136 | { |
4beb31f3 FW |
3137 | if (event->parent) |
3138 | return; | |
3139 | ||
3140 | if (has_branch_stack(event)) { | |
3141 | if (!(event->attach_state & PERF_ATTACH_TASK)) | |
3142 | atomic_dec(&per_cpu(perf_branch_stack_events, cpu)); | |
3143 | } | |
3144 | if (is_cgroup_event(event)) | |
3145 | atomic_dec(&per_cpu(perf_cgroup_events, cpu)); | |
3146 | } | |
925d519a | 3147 | |
4beb31f3 FW |
3148 | static void unaccount_event(struct perf_event *event) |
3149 | { | |
3150 | if (event->parent) | |
3151 | return; | |
3152 | ||
3153 | if (event->attach_state & PERF_ATTACH_TASK) | |
3154 | static_key_slow_dec_deferred(&perf_sched_events); | |
3155 | if (event->attr.mmap || event->attr.mmap_data) | |
3156 | atomic_dec(&nr_mmap_events); | |
3157 | if (event->attr.comm) | |
3158 | atomic_dec(&nr_comm_events); | |
3159 | if (event->attr.task) | |
3160 | atomic_dec(&nr_task_events); | |
948b26b6 FW |
3161 | if (event->attr.freq) |
3162 | atomic_dec(&nr_freq_events); | |
4beb31f3 FW |
3163 | if (is_cgroup_event(event)) |
3164 | static_key_slow_dec_deferred(&perf_sched_events); | |
3165 | if (has_branch_stack(event)) | |
3166 | static_key_slow_dec_deferred(&perf_sched_events); | |
3167 | ||
3168 | unaccount_event_cpu(event, event->cpu); | |
3169 | } | |
925d519a | 3170 | |
766d6c07 FW |
3171 | static void __free_event(struct perf_event *event) |
3172 | { | |
cdd6c482 | 3173 | if (!event->parent) { |
927c7a9e FW |
3174 | if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) |
3175 | put_callchain_buffers(); | |
f344011c | 3176 | } |
9ee318a7 | 3177 | |
766d6c07 FW |
3178 | if (event->destroy) |
3179 | event->destroy(event); | |
3180 | ||
3181 | if (event->ctx) | |
3182 | put_ctx(event->ctx); | |
3183 | ||
3184 | call_rcu(&event->rcu_head, free_event_rcu); | |
3185 | } | |
cdd6c482 | 3186 | static void free_event(struct perf_event *event) |
f1600952 | 3187 | { |
e360adbe | 3188 | irq_work_sync(&event->pending); |
925d519a | 3189 | |
4beb31f3 | 3190 | unaccount_event(event); |
9ee318a7 | 3191 | |
76369139 | 3192 | if (event->rb) { |
9bb5d40c PZ |
3193 | struct ring_buffer *rb; |
3194 | ||
3195 | /* | |
3196 | * Can happen when we close an event with re-directed output. | |
3197 | * | |
3198 | * Since we have a 0 refcount, perf_mmap_close() will skip | |
3199 | * over us; possibly making our ring_buffer_put() the last. | |
3200 | */ | |
3201 | mutex_lock(&event->mmap_mutex); | |
3202 | rb = event->rb; | |
3203 | if (rb) { | |
3204 | rcu_assign_pointer(event->rb, NULL); | |
3205 | ring_buffer_detach(event, rb); | |
3206 | ring_buffer_put(rb); /* could be last */ | |
3207 | } | |
3208 | mutex_unlock(&event->mmap_mutex); | |
a4be7c27 PZ |
3209 | } |
3210 | ||
e5d1367f SE |
3211 | if (is_cgroup_event(event)) |
3212 | perf_detach_cgroup(event); | |
3213 | ||
0c67b408 | 3214 | |
766d6c07 | 3215 | __free_event(event); |
f1600952 PZ |
3216 | } |
3217 | ||
a66a3052 | 3218 | int perf_event_release_kernel(struct perf_event *event) |
0793a61d | 3219 | { |
cdd6c482 | 3220 | struct perf_event_context *ctx = event->ctx; |
0793a61d | 3221 | |
ad3a37de | 3222 | WARN_ON_ONCE(ctx->parent_ctx); |
a0507c84 PZ |
3223 | /* |
3224 | * There are two ways this annotation is useful: | |
3225 | * | |
3226 | * 1) there is a lock recursion from perf_event_exit_task | |
3227 | * see the comment there. | |
3228 | * | |
3229 | * 2) there is a lock-inversion with mmap_sem through | |
3230 | * perf_event_read_group(), which takes faults while | |
3231 | * holding ctx->mutex, however this is called after | |
3232 | * the last filedesc died, so there is no possibility | |
3233 | * to trigger the AB-BA case. | |
3234 | */ | |
3235 | mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING); | |
050735b0 | 3236 | raw_spin_lock_irq(&ctx->lock); |
8a49542c | 3237 | perf_group_detach(event); |
050735b0 | 3238 | raw_spin_unlock_irq(&ctx->lock); |
e03a9a55 | 3239 | perf_remove_from_context(event); |
d859e29f | 3240 | mutex_unlock(&ctx->mutex); |
0793a61d | 3241 | |
cdd6c482 | 3242 | free_event(event); |
0793a61d TG |
3243 | |
3244 | return 0; | |
3245 | } | |
a66a3052 | 3246 | EXPORT_SYMBOL_GPL(perf_event_release_kernel); |
0793a61d | 3247 | |
a66a3052 PZ |
3248 | /* |
3249 | * Called when the last reference to the file is gone. | |
3250 | */ | |
a6fa941d | 3251 | static void put_event(struct perf_event *event) |
fb0459d7 | 3252 | { |
8882135b | 3253 | struct task_struct *owner; |
fb0459d7 | 3254 | |
a6fa941d AV |
3255 | if (!atomic_long_dec_and_test(&event->refcount)) |
3256 | return; | |
fb0459d7 | 3257 | |
8882135b PZ |
3258 | rcu_read_lock(); |
3259 | owner = ACCESS_ONCE(event->owner); | |
3260 | /* | |
3261 | * Matches the smp_wmb() in perf_event_exit_task(). If we observe | |
3262 | * !owner it means the list deletion is complete and we can indeed | |
3263 | * free this event, otherwise we need to serialize on | |
3264 | * owner->perf_event_mutex. | |
3265 | */ | |
3266 | smp_read_barrier_depends(); | |
3267 | if (owner) { | |
3268 | /* | |
3269 | * Since delayed_put_task_struct() also drops the last | |
3270 | * task reference we can safely take a new reference | |
3271 | * while holding the rcu_read_lock(). | |
3272 | */ | |
3273 | get_task_struct(owner); | |
3274 | } | |
3275 | rcu_read_unlock(); | |
3276 | ||
3277 | if (owner) { | |
3278 | mutex_lock(&owner->perf_event_mutex); | |
3279 | /* | |
3280 | * We have to re-check the event->owner field, if it is cleared | |
3281 | * we raced with perf_event_exit_task(), acquiring the mutex | |
3282 | * ensured they're done, and we can proceed with freeing the | |
3283 | * event. | |
3284 | */ | |
3285 | if (event->owner) | |
3286 | list_del_init(&event->owner_entry); | |
3287 | mutex_unlock(&owner->perf_event_mutex); | |
3288 | put_task_struct(owner); | |
3289 | } | |
3290 | ||
a6fa941d AV |
3291 | perf_event_release_kernel(event); |
3292 | } | |
3293 | ||
3294 | static int perf_release(struct inode *inode, struct file *file) | |
3295 | { | |
3296 | put_event(file->private_data); | |
3297 | return 0; | |
fb0459d7 | 3298 | } |
fb0459d7 | 3299 | |
59ed446f | 3300 | u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running) |
e53c0994 | 3301 | { |
cdd6c482 | 3302 | struct perf_event *child; |
e53c0994 PZ |
3303 | u64 total = 0; |
3304 | ||
59ed446f PZ |
3305 | *enabled = 0; |
3306 | *running = 0; | |
3307 | ||
6f10581a | 3308 | mutex_lock(&event->child_mutex); |
cdd6c482 | 3309 | total += perf_event_read(event); |
59ed446f PZ |
3310 | *enabled += event->total_time_enabled + |
3311 | atomic64_read(&event->child_total_time_enabled); | |
3312 | *running += event->total_time_running + | |
3313 | atomic64_read(&event->child_total_time_running); | |
3314 | ||
3315 | list_for_each_entry(child, &event->child_list, child_list) { | |
cdd6c482 | 3316 | total += perf_event_read(child); |
59ed446f PZ |
3317 | *enabled += child->total_time_enabled; |
3318 | *running += child->total_time_running; | |
3319 | } | |
6f10581a | 3320 | mutex_unlock(&event->child_mutex); |
e53c0994 PZ |
3321 | |
3322 | return total; | |
3323 | } | |
fb0459d7 | 3324 | EXPORT_SYMBOL_GPL(perf_event_read_value); |
e53c0994 | 3325 | |
cdd6c482 | 3326 | static int perf_event_read_group(struct perf_event *event, |
3dab77fb PZ |
3327 | u64 read_format, char __user *buf) |
3328 | { | |
cdd6c482 | 3329 | struct perf_event *leader = event->group_leader, *sub; |
6f10581a PZ |
3330 | int n = 0, size = 0, ret = -EFAULT; |
3331 | struct perf_event_context *ctx = leader->ctx; | |
abf4868b | 3332 | u64 values[5]; |
59ed446f | 3333 | u64 count, enabled, running; |
abf4868b | 3334 | |
6f10581a | 3335 | mutex_lock(&ctx->mutex); |
59ed446f | 3336 | count = perf_event_read_value(leader, &enabled, &running); |
3dab77fb PZ |
3337 | |
3338 | values[n++] = 1 + leader->nr_siblings; | |
59ed446f PZ |
3339 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) |
3340 | values[n++] = enabled; | |
3341 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
3342 | values[n++] = running; | |
abf4868b PZ |
3343 | values[n++] = count; |
3344 | if (read_format & PERF_FORMAT_ID) | |
3345 | values[n++] = primary_event_id(leader); | |
3dab77fb PZ |
3346 | |
3347 | size = n * sizeof(u64); | |
3348 | ||
3349 | if (copy_to_user(buf, values, size)) | |
6f10581a | 3350 | goto unlock; |
3dab77fb | 3351 | |
6f10581a | 3352 | ret = size; |
3dab77fb | 3353 | |
65abc865 | 3354 | list_for_each_entry(sub, &leader->sibling_list, group_entry) { |
abf4868b | 3355 | n = 0; |
3dab77fb | 3356 | |
59ed446f | 3357 | values[n++] = perf_event_read_value(sub, &enabled, &running); |
abf4868b PZ |
3358 | if (read_format & PERF_FORMAT_ID) |
3359 | values[n++] = primary_event_id(sub); | |
3360 | ||
3361 | size = n * sizeof(u64); | |
3362 | ||
184d3da8 | 3363 | if (copy_to_user(buf + ret, values, size)) { |
6f10581a PZ |
3364 | ret = -EFAULT; |
3365 | goto unlock; | |
3366 | } | |
abf4868b PZ |
3367 | |
3368 | ret += size; | |
3dab77fb | 3369 | } |
6f10581a PZ |
3370 | unlock: |
3371 | mutex_unlock(&ctx->mutex); | |
3dab77fb | 3372 | |
abf4868b | 3373 | return ret; |
3dab77fb PZ |
3374 | } |
3375 | ||
cdd6c482 | 3376 | static int perf_event_read_one(struct perf_event *event, |
3dab77fb PZ |
3377 | u64 read_format, char __user *buf) |
3378 | { | |
59ed446f | 3379 | u64 enabled, running; |
3dab77fb PZ |
3380 | u64 values[4]; |
3381 | int n = 0; | |
3382 | ||
59ed446f PZ |
3383 | values[n++] = perf_event_read_value(event, &enabled, &running); |
3384 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) | |
3385 | values[n++] = enabled; | |
3386 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
3387 | values[n++] = running; | |
3dab77fb | 3388 | if (read_format & PERF_FORMAT_ID) |
cdd6c482 | 3389 | values[n++] = primary_event_id(event); |
3dab77fb PZ |
3390 | |
3391 | if (copy_to_user(buf, values, n * sizeof(u64))) | |
3392 | return -EFAULT; | |
3393 | ||
3394 | return n * sizeof(u64); | |
3395 | } | |
3396 | ||
0793a61d | 3397 | /* |
cdd6c482 | 3398 | * Read the performance event - simple non blocking version for now |
0793a61d TG |
3399 | */ |
3400 | static ssize_t | |
cdd6c482 | 3401 | perf_read_hw(struct perf_event *event, char __user *buf, size_t count) |
0793a61d | 3402 | { |
cdd6c482 | 3403 | u64 read_format = event->attr.read_format; |
3dab77fb | 3404 | int ret; |
0793a61d | 3405 | |
3b6f9e5c | 3406 | /* |
cdd6c482 | 3407 | * Return end-of-file for a read on a event that is in |
3b6f9e5c PM |
3408 | * error state (i.e. because it was pinned but it couldn't be |
3409 | * scheduled on to the CPU at some point). | |
3410 | */ | |
cdd6c482 | 3411 | if (event->state == PERF_EVENT_STATE_ERROR) |
3b6f9e5c PM |
3412 | return 0; |
3413 | ||
c320c7b7 | 3414 | if (count < event->read_size) |
3dab77fb PZ |
3415 | return -ENOSPC; |
3416 | ||
cdd6c482 | 3417 | WARN_ON_ONCE(event->ctx->parent_ctx); |
3dab77fb | 3418 | if (read_format & PERF_FORMAT_GROUP) |
cdd6c482 | 3419 | ret = perf_event_read_group(event, read_format, buf); |
3dab77fb | 3420 | else |
cdd6c482 | 3421 | ret = perf_event_read_one(event, read_format, buf); |
0793a61d | 3422 | |
3dab77fb | 3423 | return ret; |
0793a61d TG |
3424 | } |
3425 | ||
0793a61d TG |
3426 | static ssize_t |
3427 | perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos) | |
3428 | { | |
cdd6c482 | 3429 | struct perf_event *event = file->private_data; |
0793a61d | 3430 | |
cdd6c482 | 3431 | return perf_read_hw(event, buf, count); |
0793a61d TG |
3432 | } |
3433 | ||
3434 | static unsigned int perf_poll(struct file *file, poll_table *wait) | |
3435 | { | |
cdd6c482 | 3436 | struct perf_event *event = file->private_data; |
76369139 | 3437 | struct ring_buffer *rb; |
c33a0bc4 | 3438 | unsigned int events = POLL_HUP; |
c7138f37 | 3439 | |
10c6db11 | 3440 | /* |
9bb5d40c PZ |
3441 | * Pin the event->rb by taking event->mmap_mutex; otherwise |
3442 | * perf_event_set_output() can swizzle our rb and make us miss wakeups. | |
10c6db11 PZ |
3443 | */ |
3444 | mutex_lock(&event->mmap_mutex); | |
9bb5d40c PZ |
3445 | rb = event->rb; |
3446 | if (rb) | |
76369139 | 3447 | events = atomic_xchg(&rb->poll, 0); |
10c6db11 PZ |
3448 | mutex_unlock(&event->mmap_mutex); |
3449 | ||
cdd6c482 | 3450 | poll_wait(file, &event->waitq, wait); |
0793a61d | 3451 | |
0793a61d TG |
3452 | return events; |
3453 | } | |
3454 | ||
cdd6c482 | 3455 | static void perf_event_reset(struct perf_event *event) |
6de6a7b9 | 3456 | { |
cdd6c482 | 3457 | (void)perf_event_read(event); |
e7850595 | 3458 | local64_set(&event->count, 0); |
cdd6c482 | 3459 | perf_event_update_userpage(event); |
3df5edad PZ |
3460 | } |
3461 | ||
c93f7669 | 3462 | /* |
cdd6c482 IM |
3463 | * Holding the top-level event's child_mutex means that any |
3464 | * descendant process that has inherited this event will block | |
3465 | * in sync_child_event if it goes to exit, thus satisfying the | |
3466 | * task existence requirements of perf_event_enable/disable. | |
c93f7669 | 3467 | */ |
cdd6c482 IM |
3468 | static void perf_event_for_each_child(struct perf_event *event, |
3469 | void (*func)(struct perf_event *)) | |
3df5edad | 3470 | { |
cdd6c482 | 3471 | struct perf_event *child; |
3df5edad | 3472 | |
cdd6c482 IM |
3473 | WARN_ON_ONCE(event->ctx->parent_ctx); |
3474 | mutex_lock(&event->child_mutex); | |
3475 | func(event); | |
3476 | list_for_each_entry(child, &event->child_list, child_list) | |
3df5edad | 3477 | func(child); |
cdd6c482 | 3478 | mutex_unlock(&event->child_mutex); |
3df5edad PZ |
3479 | } |
3480 | ||
cdd6c482 IM |
3481 | static void perf_event_for_each(struct perf_event *event, |
3482 | void (*func)(struct perf_event *)) | |
3df5edad | 3483 | { |
cdd6c482 IM |
3484 | struct perf_event_context *ctx = event->ctx; |
3485 | struct perf_event *sibling; | |
3df5edad | 3486 | |
75f937f2 PZ |
3487 | WARN_ON_ONCE(ctx->parent_ctx); |
3488 | mutex_lock(&ctx->mutex); | |
cdd6c482 | 3489 | event = event->group_leader; |
75f937f2 | 3490 | |
cdd6c482 | 3491 | perf_event_for_each_child(event, func); |
cdd6c482 | 3492 | list_for_each_entry(sibling, &event->sibling_list, group_entry) |
724b6daa | 3493 | perf_event_for_each_child(sibling, func); |
75f937f2 | 3494 | mutex_unlock(&ctx->mutex); |
6de6a7b9 PZ |
3495 | } |
3496 | ||
cdd6c482 | 3497 | static int perf_event_period(struct perf_event *event, u64 __user *arg) |
08247e31 | 3498 | { |
cdd6c482 | 3499 | struct perf_event_context *ctx = event->ctx; |
08247e31 PZ |
3500 | int ret = 0; |
3501 | u64 value; | |
3502 | ||
6c7e550f | 3503 | if (!is_sampling_event(event)) |
08247e31 PZ |
3504 | return -EINVAL; |
3505 | ||
ad0cf347 | 3506 | if (copy_from_user(&value, arg, sizeof(value))) |
08247e31 PZ |
3507 | return -EFAULT; |
3508 | ||
3509 | if (!value) | |
3510 | return -EINVAL; | |
3511 | ||
e625cce1 | 3512 | raw_spin_lock_irq(&ctx->lock); |
cdd6c482 IM |
3513 | if (event->attr.freq) { |
3514 | if (value > sysctl_perf_event_sample_rate) { | |
08247e31 PZ |
3515 | ret = -EINVAL; |
3516 | goto unlock; | |
3517 | } | |
3518 | ||
cdd6c482 | 3519 | event->attr.sample_freq = value; |
08247e31 | 3520 | } else { |
cdd6c482 IM |
3521 | event->attr.sample_period = value; |
3522 | event->hw.sample_period = value; | |
08247e31 PZ |
3523 | } |
3524 | unlock: | |
e625cce1 | 3525 | raw_spin_unlock_irq(&ctx->lock); |
08247e31 PZ |
3526 | |
3527 | return ret; | |
3528 | } | |
3529 | ||
ac9721f3 PZ |
3530 | static const struct file_operations perf_fops; |
3531 | ||
2903ff01 | 3532 | static inline int perf_fget_light(int fd, struct fd *p) |
ac9721f3 | 3533 | { |
2903ff01 AV |
3534 | struct fd f = fdget(fd); |
3535 | if (!f.file) | |
3536 | return -EBADF; | |
ac9721f3 | 3537 | |
2903ff01 AV |
3538 | if (f.file->f_op != &perf_fops) { |
3539 | fdput(f); | |
3540 | return -EBADF; | |
ac9721f3 | 3541 | } |
2903ff01 AV |
3542 | *p = f; |
3543 | return 0; | |
ac9721f3 PZ |
3544 | } |
3545 | ||
3546 | static int perf_event_set_output(struct perf_event *event, | |
3547 | struct perf_event *output_event); | |
6fb2915d | 3548 | static int perf_event_set_filter(struct perf_event *event, void __user *arg); |
a4be7c27 | 3549 | |
d859e29f PM |
3550 | static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg) |
3551 | { | |
cdd6c482 IM |
3552 | struct perf_event *event = file->private_data; |
3553 | void (*func)(struct perf_event *); | |
3df5edad | 3554 | u32 flags = arg; |
d859e29f PM |
3555 | |
3556 | switch (cmd) { | |
cdd6c482 IM |
3557 | case PERF_EVENT_IOC_ENABLE: |
3558 | func = perf_event_enable; | |
d859e29f | 3559 | break; |
cdd6c482 IM |
3560 | case PERF_EVENT_IOC_DISABLE: |
3561 | func = perf_event_disable; | |
79f14641 | 3562 | break; |
cdd6c482 IM |
3563 | case PERF_EVENT_IOC_RESET: |
3564 | func = perf_event_reset; | |
6de6a7b9 | 3565 | break; |
3df5edad | 3566 | |
cdd6c482 IM |
3567 | case PERF_EVENT_IOC_REFRESH: |
3568 | return perf_event_refresh(event, arg); | |
08247e31 | 3569 | |
cdd6c482 IM |
3570 | case PERF_EVENT_IOC_PERIOD: |
3571 | return perf_event_period(event, (u64 __user *)arg); | |
08247e31 | 3572 | |
cf4957f1 JO |
3573 | case PERF_EVENT_IOC_ID: |
3574 | { | |
3575 | u64 id = primary_event_id(event); | |
3576 | ||
3577 | if (copy_to_user((void __user *)arg, &id, sizeof(id))) | |
3578 | return -EFAULT; | |
3579 | return 0; | |
3580 | } | |
3581 | ||
cdd6c482 | 3582 | case PERF_EVENT_IOC_SET_OUTPUT: |
ac9721f3 | 3583 | { |
ac9721f3 | 3584 | int ret; |
ac9721f3 | 3585 | if (arg != -1) { |
2903ff01 AV |
3586 | struct perf_event *output_event; |
3587 | struct fd output; | |
3588 | ret = perf_fget_light(arg, &output); | |
3589 | if (ret) | |
3590 | return ret; | |
3591 | output_event = output.file->private_data; | |
3592 | ret = perf_event_set_output(event, output_event); | |
3593 | fdput(output); | |
3594 | } else { | |
3595 | ret = perf_event_set_output(event, NULL); | |
ac9721f3 | 3596 | } |
ac9721f3 PZ |
3597 | return ret; |
3598 | } | |
a4be7c27 | 3599 | |
6fb2915d LZ |
3600 | case PERF_EVENT_IOC_SET_FILTER: |
3601 | return perf_event_set_filter(event, (void __user *)arg); | |
3602 | ||
d859e29f | 3603 | default: |
3df5edad | 3604 | return -ENOTTY; |
d859e29f | 3605 | } |
3df5edad PZ |
3606 | |
3607 | if (flags & PERF_IOC_FLAG_GROUP) | |
cdd6c482 | 3608 | perf_event_for_each(event, func); |
3df5edad | 3609 | else |
cdd6c482 | 3610 | perf_event_for_each_child(event, func); |
3df5edad PZ |
3611 | |
3612 | return 0; | |
d859e29f PM |
3613 | } |
3614 | ||
cdd6c482 | 3615 | int perf_event_task_enable(void) |
771d7cde | 3616 | { |
cdd6c482 | 3617 | struct perf_event *event; |
771d7cde | 3618 | |
cdd6c482 IM |
3619 | mutex_lock(¤t->perf_event_mutex); |
3620 | list_for_each_entry(event, ¤t->perf_event_list, owner_entry) | |
3621 | perf_event_for_each_child(event, perf_event_enable); | |
3622 | mutex_unlock(¤t->perf_event_mutex); | |
771d7cde PZ |
3623 | |
3624 | return 0; | |
3625 | } | |
3626 | ||
cdd6c482 | 3627 | int perf_event_task_disable(void) |
771d7cde | 3628 | { |
cdd6c482 | 3629 | struct perf_event *event; |
771d7cde | 3630 | |
cdd6c482 IM |
3631 | mutex_lock(¤t->perf_event_mutex); |
3632 | list_for_each_entry(event, ¤t->perf_event_list, owner_entry) | |
3633 | perf_event_for_each_child(event, perf_event_disable); | |
3634 | mutex_unlock(¤t->perf_event_mutex); | |
771d7cde PZ |
3635 | |
3636 | return 0; | |
3637 | } | |
3638 | ||
cdd6c482 | 3639 | static int perf_event_index(struct perf_event *event) |
194002b2 | 3640 | { |
a4eaf7f1 PZ |
3641 | if (event->hw.state & PERF_HES_STOPPED) |
3642 | return 0; | |
3643 | ||
cdd6c482 | 3644 | if (event->state != PERF_EVENT_STATE_ACTIVE) |
194002b2 PZ |
3645 | return 0; |
3646 | ||
35edc2a5 | 3647 | return event->pmu->event_idx(event); |
194002b2 PZ |
3648 | } |
3649 | ||
c4794295 | 3650 | static void calc_timer_values(struct perf_event *event, |
e3f3541c | 3651 | u64 *now, |
7f310a5d EM |
3652 | u64 *enabled, |
3653 | u64 *running) | |
c4794295 | 3654 | { |
e3f3541c | 3655 | u64 ctx_time; |
c4794295 | 3656 | |
e3f3541c PZ |
3657 | *now = perf_clock(); |
3658 | ctx_time = event->shadow_ctx_time + *now; | |
c4794295 EM |
3659 | *enabled = ctx_time - event->tstamp_enabled; |
3660 | *running = ctx_time - event->tstamp_running; | |
3661 | } | |
3662 | ||
fa731587 PZ |
3663 | static void perf_event_init_userpage(struct perf_event *event) |
3664 | { | |
3665 | struct perf_event_mmap_page *userpg; | |
3666 | struct ring_buffer *rb; | |
3667 | ||
3668 | rcu_read_lock(); | |
3669 | rb = rcu_dereference(event->rb); | |
3670 | if (!rb) | |
3671 | goto unlock; | |
3672 | ||
3673 | userpg = rb->user_page; | |
3674 | ||
3675 | /* Allow new userspace to detect that bit 0 is deprecated */ | |
3676 | userpg->cap_bit0_is_deprecated = 1; | |
3677 | userpg->size = offsetof(struct perf_event_mmap_page, __reserved); | |
3678 | ||
3679 | unlock: | |
3680 | rcu_read_unlock(); | |
3681 | } | |
3682 | ||
c7206205 | 3683 | void __weak arch_perf_update_userpage(struct perf_event_mmap_page *userpg, u64 now) |
e3f3541c PZ |
3684 | { |
3685 | } | |
3686 | ||
38ff667b PZ |
3687 | /* |
3688 | * Callers need to ensure there can be no nesting of this function, otherwise | |
3689 | * the seqlock logic goes bad. We can not serialize this because the arch | |
3690 | * code calls this from NMI context. | |
3691 | */ | |
cdd6c482 | 3692 | void perf_event_update_userpage(struct perf_event *event) |
37d81828 | 3693 | { |
cdd6c482 | 3694 | struct perf_event_mmap_page *userpg; |
76369139 | 3695 | struct ring_buffer *rb; |
e3f3541c | 3696 | u64 enabled, running, now; |
38ff667b PZ |
3697 | |
3698 | rcu_read_lock(); | |
5ec4c599 PZ |
3699 | rb = rcu_dereference(event->rb); |
3700 | if (!rb) | |
3701 | goto unlock; | |
3702 | ||
0d641208 EM |
3703 | /* |
3704 | * compute total_time_enabled, total_time_running | |
3705 | * based on snapshot values taken when the event | |
3706 | * was last scheduled in. | |
3707 | * | |
3708 | * we cannot simply called update_context_time() | |
3709 | * because of locking issue as we can be called in | |
3710 | * NMI context | |
3711 | */ | |
e3f3541c | 3712 | calc_timer_values(event, &now, &enabled, &running); |
38ff667b | 3713 | |
76369139 | 3714 | userpg = rb->user_page; |
7b732a75 PZ |
3715 | /* |
3716 | * Disable preemption so as to not let the corresponding user-space | |
3717 | * spin too long if we get preempted. | |
3718 | */ | |
3719 | preempt_disable(); | |
37d81828 | 3720 | ++userpg->lock; |
92f22a38 | 3721 | barrier(); |
cdd6c482 | 3722 | userpg->index = perf_event_index(event); |
b5e58793 | 3723 | userpg->offset = perf_event_count(event); |
365a4038 | 3724 | if (userpg->index) |
e7850595 | 3725 | userpg->offset -= local64_read(&event->hw.prev_count); |
7b732a75 | 3726 | |
0d641208 | 3727 | userpg->time_enabled = enabled + |
cdd6c482 | 3728 | atomic64_read(&event->child_total_time_enabled); |
7f8b4e4e | 3729 | |
0d641208 | 3730 | userpg->time_running = running + |
cdd6c482 | 3731 | atomic64_read(&event->child_total_time_running); |
7f8b4e4e | 3732 | |
c7206205 | 3733 | arch_perf_update_userpage(userpg, now); |
e3f3541c | 3734 | |
92f22a38 | 3735 | barrier(); |
37d81828 | 3736 | ++userpg->lock; |
7b732a75 | 3737 | preempt_enable(); |
38ff667b | 3738 | unlock: |
7b732a75 | 3739 | rcu_read_unlock(); |
37d81828 PM |
3740 | } |
3741 | ||
906010b2 PZ |
3742 | static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf) |
3743 | { | |
3744 | struct perf_event *event = vma->vm_file->private_data; | |
76369139 | 3745 | struct ring_buffer *rb; |
906010b2 PZ |
3746 | int ret = VM_FAULT_SIGBUS; |
3747 | ||
3748 | if (vmf->flags & FAULT_FLAG_MKWRITE) { | |
3749 | if (vmf->pgoff == 0) | |
3750 | ret = 0; | |
3751 | return ret; | |
3752 | } | |
3753 | ||
3754 | rcu_read_lock(); | |
76369139 FW |
3755 | rb = rcu_dereference(event->rb); |
3756 | if (!rb) | |
906010b2 PZ |
3757 | goto unlock; |
3758 | ||
3759 | if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE)) | |
3760 | goto unlock; | |
3761 | ||
76369139 | 3762 | vmf->page = perf_mmap_to_page(rb, vmf->pgoff); |
906010b2 PZ |
3763 | if (!vmf->page) |
3764 | goto unlock; | |
3765 | ||
3766 | get_page(vmf->page); | |
3767 | vmf->page->mapping = vma->vm_file->f_mapping; | |
3768 | vmf->page->index = vmf->pgoff; | |
3769 | ||
3770 | ret = 0; | |
3771 | unlock: | |
3772 | rcu_read_unlock(); | |
3773 | ||
3774 | return ret; | |
3775 | } | |
3776 | ||
10c6db11 PZ |
3777 | static void ring_buffer_attach(struct perf_event *event, |
3778 | struct ring_buffer *rb) | |
3779 | { | |
3780 | unsigned long flags; | |
3781 | ||
3782 | if (!list_empty(&event->rb_entry)) | |
3783 | return; | |
3784 | ||
3785 | spin_lock_irqsave(&rb->event_lock, flags); | |
9bb5d40c PZ |
3786 | if (list_empty(&event->rb_entry)) |
3787 | list_add(&event->rb_entry, &rb->event_list); | |
10c6db11 PZ |
3788 | spin_unlock_irqrestore(&rb->event_lock, flags); |
3789 | } | |
3790 | ||
9bb5d40c | 3791 | static void ring_buffer_detach(struct perf_event *event, struct ring_buffer *rb) |
10c6db11 PZ |
3792 | { |
3793 | unsigned long flags; | |
3794 | ||
3795 | if (list_empty(&event->rb_entry)) | |
3796 | return; | |
3797 | ||
3798 | spin_lock_irqsave(&rb->event_lock, flags); | |
3799 | list_del_init(&event->rb_entry); | |
3800 | wake_up_all(&event->waitq); | |
3801 | spin_unlock_irqrestore(&rb->event_lock, flags); | |
3802 | } | |
3803 | ||
3804 | static void ring_buffer_wakeup(struct perf_event *event) | |
3805 | { | |
3806 | struct ring_buffer *rb; | |
3807 | ||
3808 | rcu_read_lock(); | |
3809 | rb = rcu_dereference(event->rb); | |
9bb5d40c PZ |
3810 | if (rb) { |
3811 | list_for_each_entry_rcu(event, &rb->event_list, rb_entry) | |
3812 | wake_up_all(&event->waitq); | |
3813 | } | |
10c6db11 PZ |
3814 | rcu_read_unlock(); |
3815 | } | |
3816 | ||
76369139 | 3817 | static void rb_free_rcu(struct rcu_head *rcu_head) |
906010b2 | 3818 | { |
76369139 | 3819 | struct ring_buffer *rb; |
906010b2 | 3820 | |
76369139 FW |
3821 | rb = container_of(rcu_head, struct ring_buffer, rcu_head); |
3822 | rb_free(rb); | |
7b732a75 PZ |
3823 | } |
3824 | ||
76369139 | 3825 | static struct ring_buffer *ring_buffer_get(struct perf_event *event) |
7b732a75 | 3826 | { |
76369139 | 3827 | struct ring_buffer *rb; |
7b732a75 | 3828 | |
ac9721f3 | 3829 | rcu_read_lock(); |
76369139 FW |
3830 | rb = rcu_dereference(event->rb); |
3831 | if (rb) { | |
3832 | if (!atomic_inc_not_zero(&rb->refcount)) | |
3833 | rb = NULL; | |
ac9721f3 PZ |
3834 | } |
3835 | rcu_read_unlock(); | |
3836 | ||
76369139 | 3837 | return rb; |
ac9721f3 PZ |
3838 | } |
3839 | ||
76369139 | 3840 | static void ring_buffer_put(struct ring_buffer *rb) |
ac9721f3 | 3841 | { |
76369139 | 3842 | if (!atomic_dec_and_test(&rb->refcount)) |
ac9721f3 | 3843 | return; |
7b732a75 | 3844 | |
9bb5d40c | 3845 | WARN_ON_ONCE(!list_empty(&rb->event_list)); |
10c6db11 | 3846 | |
76369139 | 3847 | call_rcu(&rb->rcu_head, rb_free_rcu); |
7b732a75 PZ |
3848 | } |
3849 | ||
3850 | static void perf_mmap_open(struct vm_area_struct *vma) | |
3851 | { | |
cdd6c482 | 3852 | struct perf_event *event = vma->vm_file->private_data; |
7b732a75 | 3853 | |
cdd6c482 | 3854 | atomic_inc(&event->mmap_count); |
9bb5d40c | 3855 | atomic_inc(&event->rb->mmap_count); |
7b732a75 PZ |
3856 | } |
3857 | ||
9bb5d40c PZ |
3858 | /* |
3859 | * A buffer can be mmap()ed multiple times; either directly through the same | |
3860 | * event, or through other events by use of perf_event_set_output(). | |
3861 | * | |
3862 | * In order to undo the VM accounting done by perf_mmap() we need to destroy | |
3863 | * the buffer here, where we still have a VM context. This means we need | |
3864 | * to detach all events redirecting to us. | |
3865 | */ | |
7b732a75 PZ |
3866 | static void perf_mmap_close(struct vm_area_struct *vma) |
3867 | { | |
cdd6c482 | 3868 | struct perf_event *event = vma->vm_file->private_data; |
7b732a75 | 3869 | |
9bb5d40c PZ |
3870 | struct ring_buffer *rb = event->rb; |
3871 | struct user_struct *mmap_user = rb->mmap_user; | |
3872 | int mmap_locked = rb->mmap_locked; | |
3873 | unsigned long size = perf_data_size(rb); | |
789f90fc | 3874 | |
9bb5d40c PZ |
3875 | atomic_dec(&rb->mmap_count); |
3876 | ||
3877 | if (!atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) | |
3878 | return; | |
3879 | ||
3880 | /* Detach current event from the buffer. */ | |
3881 | rcu_assign_pointer(event->rb, NULL); | |
3882 | ring_buffer_detach(event, rb); | |
3883 | mutex_unlock(&event->mmap_mutex); | |
3884 | ||
3885 | /* If there's still other mmap()s of this buffer, we're done. */ | |
3886 | if (atomic_read(&rb->mmap_count)) { | |
3887 | ring_buffer_put(rb); /* can't be last */ | |
3888 | return; | |
3889 | } | |
ac9721f3 | 3890 | |
9bb5d40c PZ |
3891 | /* |
3892 | * No other mmap()s, detach from all other events that might redirect | |
3893 | * into the now unreachable buffer. Somewhat complicated by the | |
3894 | * fact that rb::event_lock otherwise nests inside mmap_mutex. | |
3895 | */ | |
3896 | again: | |
3897 | rcu_read_lock(); | |
3898 | list_for_each_entry_rcu(event, &rb->event_list, rb_entry) { | |
3899 | if (!atomic_long_inc_not_zero(&event->refcount)) { | |
3900 | /* | |
3901 | * This event is en-route to free_event() which will | |
3902 | * detach it and remove it from the list. | |
3903 | */ | |
3904 | continue; | |
3905 | } | |
3906 | rcu_read_unlock(); | |
789f90fc | 3907 | |
9bb5d40c PZ |
3908 | mutex_lock(&event->mmap_mutex); |
3909 | /* | |
3910 | * Check we didn't race with perf_event_set_output() which can | |
3911 | * swizzle the rb from under us while we were waiting to | |
3912 | * acquire mmap_mutex. | |
3913 | * | |
3914 | * If we find a different rb; ignore this event, a next | |
3915 | * iteration will no longer find it on the list. We have to | |
3916 | * still restart the iteration to make sure we're not now | |
3917 | * iterating the wrong list. | |
3918 | */ | |
3919 | if (event->rb == rb) { | |
3920 | rcu_assign_pointer(event->rb, NULL); | |
3921 | ring_buffer_detach(event, rb); | |
3922 | ring_buffer_put(rb); /* can't be last, we still have one */ | |
26cb63ad | 3923 | } |
cdd6c482 | 3924 | mutex_unlock(&event->mmap_mutex); |
9bb5d40c | 3925 | put_event(event); |
ac9721f3 | 3926 | |
9bb5d40c PZ |
3927 | /* |
3928 | * Restart the iteration; either we're on the wrong list or | |
3929 | * destroyed its integrity by doing a deletion. | |
3930 | */ | |
3931 | goto again; | |
7b732a75 | 3932 | } |
9bb5d40c PZ |
3933 | rcu_read_unlock(); |
3934 | ||
3935 | /* | |
3936 | * It could be there's still a few 0-ref events on the list; they'll | |
3937 | * get cleaned up by free_event() -- they'll also still have their | |
3938 | * ref on the rb and will free it whenever they are done with it. | |
3939 | * | |
3940 | * Aside from that, this buffer is 'fully' detached and unmapped, | |
3941 | * undo the VM accounting. | |
3942 | */ | |
3943 | ||
3944 | atomic_long_sub((size >> PAGE_SHIFT) + 1, &mmap_user->locked_vm); | |
3945 | vma->vm_mm->pinned_vm -= mmap_locked; | |
3946 | free_uid(mmap_user); | |
3947 | ||
3948 | ring_buffer_put(rb); /* could be last */ | |
37d81828 PM |
3949 | } |
3950 | ||
f0f37e2f | 3951 | static const struct vm_operations_struct perf_mmap_vmops = { |
43a21ea8 PZ |
3952 | .open = perf_mmap_open, |
3953 | .close = perf_mmap_close, | |
3954 | .fault = perf_mmap_fault, | |
3955 | .page_mkwrite = perf_mmap_fault, | |
37d81828 PM |
3956 | }; |
3957 | ||
3958 | static int perf_mmap(struct file *file, struct vm_area_struct *vma) | |
3959 | { | |
cdd6c482 | 3960 | struct perf_event *event = file->private_data; |
22a4f650 | 3961 | unsigned long user_locked, user_lock_limit; |
789f90fc | 3962 | struct user_struct *user = current_user(); |
22a4f650 | 3963 | unsigned long locked, lock_limit; |
76369139 | 3964 | struct ring_buffer *rb; |
7b732a75 PZ |
3965 | unsigned long vma_size; |
3966 | unsigned long nr_pages; | |
789f90fc | 3967 | long user_extra, extra; |
d57e34fd | 3968 | int ret = 0, flags = 0; |
37d81828 | 3969 | |
c7920614 PZ |
3970 | /* |
3971 | * Don't allow mmap() of inherited per-task counters. This would | |
3972 | * create a performance issue due to all children writing to the | |
76369139 | 3973 | * same rb. |
c7920614 PZ |
3974 | */ |
3975 | if (event->cpu == -1 && event->attr.inherit) | |
3976 | return -EINVAL; | |
3977 | ||
43a21ea8 | 3978 | if (!(vma->vm_flags & VM_SHARED)) |
37d81828 | 3979 | return -EINVAL; |
7b732a75 PZ |
3980 | |
3981 | vma_size = vma->vm_end - vma->vm_start; | |
3982 | nr_pages = (vma_size / PAGE_SIZE) - 1; | |
3983 | ||
7730d865 | 3984 | /* |
76369139 | 3985 | * If we have rb pages ensure they're a power-of-two number, so we |
7730d865 PZ |
3986 | * can do bitmasks instead of modulo. |
3987 | */ | |
3988 | if (nr_pages != 0 && !is_power_of_2(nr_pages)) | |
37d81828 PM |
3989 | return -EINVAL; |
3990 | ||
7b732a75 | 3991 | if (vma_size != PAGE_SIZE * (1 + nr_pages)) |
37d81828 PM |
3992 | return -EINVAL; |
3993 | ||
7b732a75 PZ |
3994 | if (vma->vm_pgoff != 0) |
3995 | return -EINVAL; | |
37d81828 | 3996 | |
cdd6c482 | 3997 | WARN_ON_ONCE(event->ctx->parent_ctx); |
9bb5d40c | 3998 | again: |
cdd6c482 | 3999 | mutex_lock(&event->mmap_mutex); |
76369139 | 4000 | if (event->rb) { |
9bb5d40c | 4001 | if (event->rb->nr_pages != nr_pages) { |
ebb3c4c4 | 4002 | ret = -EINVAL; |
9bb5d40c PZ |
4003 | goto unlock; |
4004 | } | |
4005 | ||
4006 | if (!atomic_inc_not_zero(&event->rb->mmap_count)) { | |
4007 | /* | |
4008 | * Raced against perf_mmap_close() through | |
4009 | * perf_event_set_output(). Try again, hope for better | |
4010 | * luck. | |
4011 | */ | |
4012 | mutex_unlock(&event->mmap_mutex); | |
4013 | goto again; | |
4014 | } | |
4015 | ||
ebb3c4c4 PZ |
4016 | goto unlock; |
4017 | } | |
4018 | ||
789f90fc | 4019 | user_extra = nr_pages + 1; |
cdd6c482 | 4020 | user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10); |
a3862d3f IM |
4021 | |
4022 | /* | |
4023 | * Increase the limit linearly with more CPUs: | |
4024 | */ | |
4025 | user_lock_limit *= num_online_cpus(); | |
4026 | ||
789f90fc | 4027 | user_locked = atomic_long_read(&user->locked_vm) + user_extra; |
c5078f78 | 4028 | |
789f90fc PZ |
4029 | extra = 0; |
4030 | if (user_locked > user_lock_limit) | |
4031 | extra = user_locked - user_lock_limit; | |
7b732a75 | 4032 | |
78d7d407 | 4033 | lock_limit = rlimit(RLIMIT_MEMLOCK); |
7b732a75 | 4034 | lock_limit >>= PAGE_SHIFT; |
bc3e53f6 | 4035 | locked = vma->vm_mm->pinned_vm + extra; |
7b732a75 | 4036 | |
459ec28a IM |
4037 | if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() && |
4038 | !capable(CAP_IPC_LOCK)) { | |
ebb3c4c4 PZ |
4039 | ret = -EPERM; |
4040 | goto unlock; | |
4041 | } | |
7b732a75 | 4042 | |
76369139 | 4043 | WARN_ON(event->rb); |
906010b2 | 4044 | |
d57e34fd | 4045 | if (vma->vm_flags & VM_WRITE) |
76369139 | 4046 | flags |= RING_BUFFER_WRITABLE; |
d57e34fd | 4047 | |
4ec8363d VW |
4048 | rb = rb_alloc(nr_pages, |
4049 | event->attr.watermark ? event->attr.wakeup_watermark : 0, | |
4050 | event->cpu, flags); | |
4051 | ||
76369139 | 4052 | if (!rb) { |
ac9721f3 | 4053 | ret = -ENOMEM; |
ebb3c4c4 | 4054 | goto unlock; |
ac9721f3 | 4055 | } |
26cb63ad | 4056 | |
9bb5d40c | 4057 | atomic_set(&rb->mmap_count, 1); |
26cb63ad PZ |
4058 | rb->mmap_locked = extra; |
4059 | rb->mmap_user = get_current_user(); | |
43a21ea8 | 4060 | |
ac9721f3 | 4061 | atomic_long_add(user_extra, &user->locked_vm); |
26cb63ad PZ |
4062 | vma->vm_mm->pinned_vm += extra; |
4063 | ||
9bb5d40c | 4064 | ring_buffer_attach(event, rb); |
26cb63ad | 4065 | rcu_assign_pointer(event->rb, rb); |
ac9721f3 | 4066 | |
fa731587 | 4067 | perf_event_init_userpage(event); |
9a0f05cb PZ |
4068 | perf_event_update_userpage(event); |
4069 | ||
ebb3c4c4 | 4070 | unlock: |
ac9721f3 PZ |
4071 | if (!ret) |
4072 | atomic_inc(&event->mmap_count); | |
cdd6c482 | 4073 | mutex_unlock(&event->mmap_mutex); |
37d81828 | 4074 | |
9bb5d40c PZ |
4075 | /* |
4076 | * Since pinned accounting is per vm we cannot allow fork() to copy our | |
4077 | * vma. | |
4078 | */ | |
26cb63ad | 4079 | vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND | VM_DONTDUMP; |
37d81828 | 4080 | vma->vm_ops = &perf_mmap_vmops; |
7b732a75 PZ |
4081 | |
4082 | return ret; | |
37d81828 PM |
4083 | } |
4084 | ||
3c446b3d PZ |
4085 | static int perf_fasync(int fd, struct file *filp, int on) |
4086 | { | |
496ad9aa | 4087 | struct inode *inode = file_inode(filp); |
cdd6c482 | 4088 | struct perf_event *event = filp->private_data; |
3c446b3d PZ |
4089 | int retval; |
4090 | ||
4091 | mutex_lock(&inode->i_mutex); | |
cdd6c482 | 4092 | retval = fasync_helper(fd, filp, on, &event->fasync); |
3c446b3d PZ |
4093 | mutex_unlock(&inode->i_mutex); |
4094 | ||
4095 | if (retval < 0) | |
4096 | return retval; | |
4097 | ||
4098 | return 0; | |
4099 | } | |
4100 | ||
0793a61d | 4101 | static const struct file_operations perf_fops = { |
3326c1ce | 4102 | .llseek = no_llseek, |
0793a61d TG |
4103 | .release = perf_release, |
4104 | .read = perf_read, | |
4105 | .poll = perf_poll, | |
d859e29f PM |
4106 | .unlocked_ioctl = perf_ioctl, |
4107 | .compat_ioctl = perf_ioctl, | |
37d81828 | 4108 | .mmap = perf_mmap, |
3c446b3d | 4109 | .fasync = perf_fasync, |
0793a61d TG |
4110 | }; |
4111 | ||
925d519a | 4112 | /* |
cdd6c482 | 4113 | * Perf event wakeup |
925d519a PZ |
4114 | * |
4115 | * If there's data, ensure we set the poll() state and publish everything | |
4116 | * to user-space before waking everybody up. | |
4117 | */ | |
4118 | ||
cdd6c482 | 4119 | void perf_event_wakeup(struct perf_event *event) |
925d519a | 4120 | { |
10c6db11 | 4121 | ring_buffer_wakeup(event); |
4c9e2542 | 4122 | |
cdd6c482 IM |
4123 | if (event->pending_kill) { |
4124 | kill_fasync(&event->fasync, SIGIO, event->pending_kill); | |
4125 | event->pending_kill = 0; | |
4c9e2542 | 4126 | } |
925d519a PZ |
4127 | } |
4128 | ||
e360adbe | 4129 | static void perf_pending_event(struct irq_work *entry) |
79f14641 | 4130 | { |
cdd6c482 IM |
4131 | struct perf_event *event = container_of(entry, |
4132 | struct perf_event, pending); | |
79f14641 | 4133 | |
cdd6c482 IM |
4134 | if (event->pending_disable) { |
4135 | event->pending_disable = 0; | |
4136 | __perf_event_disable(event); | |
79f14641 PZ |
4137 | } |
4138 | ||
cdd6c482 IM |
4139 | if (event->pending_wakeup) { |
4140 | event->pending_wakeup = 0; | |
4141 | perf_event_wakeup(event); | |
79f14641 PZ |
4142 | } |
4143 | } | |
4144 | ||
39447b38 ZY |
4145 | /* |
4146 | * We assume there is only KVM supporting the callbacks. | |
4147 | * Later on, we might change it to a list if there is | |
4148 | * another virtualization implementation supporting the callbacks. | |
4149 | */ | |
4150 | struct perf_guest_info_callbacks *perf_guest_cbs; | |
4151 | ||
4152 | int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs) | |
4153 | { | |
4154 | perf_guest_cbs = cbs; | |
4155 | return 0; | |
4156 | } | |
4157 | EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks); | |
4158 | ||
4159 | int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs) | |
4160 | { | |
4161 | perf_guest_cbs = NULL; | |
4162 | return 0; | |
4163 | } | |
4164 | EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks); | |
4165 | ||
4018994f JO |
4166 | static void |
4167 | perf_output_sample_regs(struct perf_output_handle *handle, | |
4168 | struct pt_regs *regs, u64 mask) | |
4169 | { | |
4170 | int bit; | |
4171 | ||
4172 | for_each_set_bit(bit, (const unsigned long *) &mask, | |
4173 | sizeof(mask) * BITS_PER_BYTE) { | |
4174 | u64 val; | |
4175 | ||
4176 | val = perf_reg_value(regs, bit); | |
4177 | perf_output_put(handle, val); | |
4178 | } | |
4179 | } | |
4180 | ||
4181 | static void perf_sample_regs_user(struct perf_regs_user *regs_user, | |
4182 | struct pt_regs *regs) | |
4183 | { | |
4184 | if (!user_mode(regs)) { | |
4185 | if (current->mm) | |
4186 | regs = task_pt_regs(current); | |
4187 | else | |
4188 | regs = NULL; | |
4189 | } | |
4190 | ||
4191 | if (regs) { | |
4192 | regs_user->regs = regs; | |
4193 | regs_user->abi = perf_reg_abi(current); | |
4194 | } | |
4195 | } | |
4196 | ||
c5ebcedb JO |
4197 | /* |
4198 | * Get remaining task size from user stack pointer. | |
4199 | * | |
4200 | * It'd be better to take stack vma map and limit this more | |
4201 | * precisly, but there's no way to get it safely under interrupt, | |
4202 | * so using TASK_SIZE as limit. | |
4203 | */ | |
4204 | static u64 perf_ustack_task_size(struct pt_regs *regs) | |
4205 | { | |
4206 | unsigned long addr = perf_user_stack_pointer(regs); | |
4207 | ||
4208 | if (!addr || addr >= TASK_SIZE) | |
4209 | return 0; | |
4210 | ||
4211 | return TASK_SIZE - addr; | |
4212 | } | |
4213 | ||
4214 | static u16 | |
4215 | perf_sample_ustack_size(u16 stack_size, u16 header_size, | |
4216 | struct pt_regs *regs) | |
4217 | { | |
4218 | u64 task_size; | |
4219 | ||
4220 | /* No regs, no stack pointer, no dump. */ | |
4221 | if (!regs) | |
4222 | return 0; | |
4223 | ||
4224 | /* | |
4225 | * Check if we fit in with the requested stack size into the: | |
4226 | * - TASK_SIZE | |
4227 | * If we don't, we limit the size to the TASK_SIZE. | |
4228 | * | |
4229 | * - remaining sample size | |
4230 | * If we don't, we customize the stack size to | |
4231 | * fit in to the remaining sample size. | |
4232 | */ | |
4233 | ||
4234 | task_size = min((u64) USHRT_MAX, perf_ustack_task_size(regs)); | |
4235 | stack_size = min(stack_size, (u16) task_size); | |
4236 | ||
4237 | /* Current header size plus static size and dynamic size. */ | |
4238 | header_size += 2 * sizeof(u64); | |
4239 | ||
4240 | /* Do we fit in with the current stack dump size? */ | |
4241 | if ((u16) (header_size + stack_size) < header_size) { | |
4242 | /* | |
4243 | * If we overflow the maximum size for the sample, | |
4244 | * we customize the stack dump size to fit in. | |
4245 | */ | |
4246 | stack_size = USHRT_MAX - header_size - sizeof(u64); | |
4247 | stack_size = round_up(stack_size, sizeof(u64)); | |
4248 | } | |
4249 | ||
4250 | return stack_size; | |
4251 | } | |
4252 | ||
4253 | static void | |
4254 | perf_output_sample_ustack(struct perf_output_handle *handle, u64 dump_size, | |
4255 | struct pt_regs *regs) | |
4256 | { | |
4257 | /* Case of a kernel thread, nothing to dump */ | |
4258 | if (!regs) { | |
4259 | u64 size = 0; | |
4260 | perf_output_put(handle, size); | |
4261 | } else { | |
4262 | unsigned long sp; | |
4263 | unsigned int rem; | |
4264 | u64 dyn_size; | |
4265 | ||
4266 | /* | |
4267 | * We dump: | |
4268 | * static size | |
4269 | * - the size requested by user or the best one we can fit | |
4270 | * in to the sample max size | |
4271 | * data | |
4272 | * - user stack dump data | |
4273 | * dynamic size | |
4274 | * - the actual dumped size | |
4275 | */ | |
4276 | ||
4277 | /* Static size. */ | |
4278 | perf_output_put(handle, dump_size); | |
4279 | ||
4280 | /* Data. */ | |
4281 | sp = perf_user_stack_pointer(regs); | |
4282 | rem = __output_copy_user(handle, (void *) sp, dump_size); | |
4283 | dyn_size = dump_size - rem; | |
4284 | ||
4285 | perf_output_skip(handle, rem); | |
4286 | ||
4287 | /* Dynamic size. */ | |
4288 | perf_output_put(handle, dyn_size); | |
4289 | } | |
4290 | } | |
4291 | ||
c980d109 ACM |
4292 | static void __perf_event_header__init_id(struct perf_event_header *header, |
4293 | struct perf_sample_data *data, | |
4294 | struct perf_event *event) | |
6844c09d ACM |
4295 | { |
4296 | u64 sample_type = event->attr.sample_type; | |
4297 | ||
4298 | data->type = sample_type; | |
4299 | header->size += event->id_header_size; | |
4300 | ||
4301 | if (sample_type & PERF_SAMPLE_TID) { | |
4302 | /* namespace issues */ | |
4303 | data->tid_entry.pid = perf_event_pid(event, current); | |
4304 | data->tid_entry.tid = perf_event_tid(event, current); | |
4305 | } | |
4306 | ||
4307 | if (sample_type & PERF_SAMPLE_TIME) | |
4308 | data->time = perf_clock(); | |
4309 | ||
ff3d527c | 4310 | if (sample_type & (PERF_SAMPLE_ID | PERF_SAMPLE_IDENTIFIER)) |
6844c09d ACM |
4311 | data->id = primary_event_id(event); |
4312 | ||
4313 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
4314 | data->stream_id = event->id; | |
4315 | ||
4316 | if (sample_type & PERF_SAMPLE_CPU) { | |
4317 | data->cpu_entry.cpu = raw_smp_processor_id(); | |
4318 | data->cpu_entry.reserved = 0; | |
4319 | } | |
4320 | } | |
4321 | ||
76369139 FW |
4322 | void perf_event_header__init_id(struct perf_event_header *header, |
4323 | struct perf_sample_data *data, | |
4324 | struct perf_event *event) | |
c980d109 ACM |
4325 | { |
4326 | if (event->attr.sample_id_all) | |
4327 | __perf_event_header__init_id(header, data, event); | |
4328 | } | |
4329 | ||
4330 | static void __perf_event__output_id_sample(struct perf_output_handle *handle, | |
4331 | struct perf_sample_data *data) | |
4332 | { | |
4333 | u64 sample_type = data->type; | |
4334 | ||
4335 | if (sample_type & PERF_SAMPLE_TID) | |
4336 | perf_output_put(handle, data->tid_entry); | |
4337 | ||
4338 | if (sample_type & PERF_SAMPLE_TIME) | |
4339 | perf_output_put(handle, data->time); | |
4340 | ||
4341 | if (sample_type & PERF_SAMPLE_ID) | |
4342 | perf_output_put(handle, data->id); | |
4343 | ||
4344 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
4345 | perf_output_put(handle, data->stream_id); | |
4346 | ||
4347 | if (sample_type & PERF_SAMPLE_CPU) | |
4348 | perf_output_put(handle, data->cpu_entry); | |
ff3d527c AH |
4349 | |
4350 | if (sample_type & PERF_SAMPLE_IDENTIFIER) | |
4351 | perf_output_put(handle, data->id); | |
c980d109 ACM |
4352 | } |
4353 | ||
76369139 FW |
4354 | void perf_event__output_id_sample(struct perf_event *event, |
4355 | struct perf_output_handle *handle, | |
4356 | struct perf_sample_data *sample) | |
c980d109 ACM |
4357 | { |
4358 | if (event->attr.sample_id_all) | |
4359 | __perf_event__output_id_sample(handle, sample); | |
4360 | } | |
4361 | ||
3dab77fb | 4362 | static void perf_output_read_one(struct perf_output_handle *handle, |
eed01528 SE |
4363 | struct perf_event *event, |
4364 | u64 enabled, u64 running) | |
3dab77fb | 4365 | { |
cdd6c482 | 4366 | u64 read_format = event->attr.read_format; |
3dab77fb PZ |
4367 | u64 values[4]; |
4368 | int n = 0; | |
4369 | ||
b5e58793 | 4370 | values[n++] = perf_event_count(event); |
3dab77fb | 4371 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) { |
eed01528 | 4372 | values[n++] = enabled + |
cdd6c482 | 4373 | atomic64_read(&event->child_total_time_enabled); |
3dab77fb PZ |
4374 | } |
4375 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) { | |
eed01528 | 4376 | values[n++] = running + |
cdd6c482 | 4377 | atomic64_read(&event->child_total_time_running); |
3dab77fb PZ |
4378 | } |
4379 | if (read_format & PERF_FORMAT_ID) | |
cdd6c482 | 4380 | values[n++] = primary_event_id(event); |
3dab77fb | 4381 | |
76369139 | 4382 | __output_copy(handle, values, n * sizeof(u64)); |
3dab77fb PZ |
4383 | } |
4384 | ||
4385 | /* | |
cdd6c482 | 4386 | * XXX PERF_FORMAT_GROUP vs inherited events seems difficult. |
3dab77fb PZ |
4387 | */ |
4388 | static void perf_output_read_group(struct perf_output_handle *handle, | |
eed01528 SE |
4389 | struct perf_event *event, |
4390 | u64 enabled, u64 running) | |
3dab77fb | 4391 | { |
cdd6c482 IM |
4392 | struct perf_event *leader = event->group_leader, *sub; |
4393 | u64 read_format = event->attr.read_format; | |
3dab77fb PZ |
4394 | u64 values[5]; |
4395 | int n = 0; | |
4396 | ||
4397 | values[n++] = 1 + leader->nr_siblings; | |
4398 | ||
4399 | if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) | |
eed01528 | 4400 | values[n++] = enabled; |
3dab77fb PZ |
4401 | |
4402 | if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) | |
eed01528 | 4403 | values[n++] = running; |
3dab77fb | 4404 | |
cdd6c482 | 4405 | if (leader != event) |
3dab77fb PZ |
4406 | leader->pmu->read(leader); |
4407 | ||
b5e58793 | 4408 | values[n++] = perf_event_count(leader); |
3dab77fb | 4409 | if (read_format & PERF_FORMAT_ID) |
cdd6c482 | 4410 | values[n++] = primary_event_id(leader); |
3dab77fb | 4411 | |
76369139 | 4412 | __output_copy(handle, values, n * sizeof(u64)); |
3dab77fb | 4413 | |
65abc865 | 4414 | list_for_each_entry(sub, &leader->sibling_list, group_entry) { |
3dab77fb PZ |
4415 | n = 0; |
4416 | ||
6f5ab001 JO |
4417 | if ((sub != event) && |
4418 | (sub->state == PERF_EVENT_STATE_ACTIVE)) | |
3dab77fb PZ |
4419 | sub->pmu->read(sub); |
4420 | ||
b5e58793 | 4421 | values[n++] = perf_event_count(sub); |
3dab77fb | 4422 | if (read_format & PERF_FORMAT_ID) |
cdd6c482 | 4423 | values[n++] = primary_event_id(sub); |
3dab77fb | 4424 | |
76369139 | 4425 | __output_copy(handle, values, n * sizeof(u64)); |
3dab77fb PZ |
4426 | } |
4427 | } | |
4428 | ||
eed01528 SE |
4429 | #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\ |
4430 | PERF_FORMAT_TOTAL_TIME_RUNNING) | |
4431 | ||
3dab77fb | 4432 | static void perf_output_read(struct perf_output_handle *handle, |
cdd6c482 | 4433 | struct perf_event *event) |
3dab77fb | 4434 | { |
e3f3541c | 4435 | u64 enabled = 0, running = 0, now; |
eed01528 SE |
4436 | u64 read_format = event->attr.read_format; |
4437 | ||
4438 | /* | |
4439 | * compute total_time_enabled, total_time_running | |
4440 | * based on snapshot values taken when the event | |
4441 | * was last scheduled in. | |
4442 | * | |
4443 | * we cannot simply called update_context_time() | |
4444 | * because of locking issue as we are called in | |
4445 | * NMI context | |
4446 | */ | |
c4794295 | 4447 | if (read_format & PERF_FORMAT_TOTAL_TIMES) |
e3f3541c | 4448 | calc_timer_values(event, &now, &enabled, &running); |
eed01528 | 4449 | |
cdd6c482 | 4450 | if (event->attr.read_format & PERF_FORMAT_GROUP) |
eed01528 | 4451 | perf_output_read_group(handle, event, enabled, running); |
3dab77fb | 4452 | else |
eed01528 | 4453 | perf_output_read_one(handle, event, enabled, running); |
3dab77fb PZ |
4454 | } |
4455 | ||
5622f295 MM |
4456 | void perf_output_sample(struct perf_output_handle *handle, |
4457 | struct perf_event_header *header, | |
4458 | struct perf_sample_data *data, | |
cdd6c482 | 4459 | struct perf_event *event) |
5622f295 MM |
4460 | { |
4461 | u64 sample_type = data->type; | |
4462 | ||
4463 | perf_output_put(handle, *header); | |
4464 | ||
ff3d527c AH |
4465 | if (sample_type & PERF_SAMPLE_IDENTIFIER) |
4466 | perf_output_put(handle, data->id); | |
4467 | ||
5622f295 MM |
4468 | if (sample_type & PERF_SAMPLE_IP) |
4469 | perf_output_put(handle, data->ip); | |
4470 | ||
4471 | if (sample_type & PERF_SAMPLE_TID) | |
4472 | perf_output_put(handle, data->tid_entry); | |
4473 | ||
4474 | if (sample_type & PERF_SAMPLE_TIME) | |
4475 | perf_output_put(handle, data->time); | |
4476 | ||
4477 | if (sample_type & PERF_SAMPLE_ADDR) | |
4478 | perf_output_put(handle, data->addr); | |
4479 | ||
4480 | if (sample_type & PERF_SAMPLE_ID) | |
4481 | perf_output_put(handle, data->id); | |
4482 | ||
4483 | if (sample_type & PERF_SAMPLE_STREAM_ID) | |
4484 | perf_output_put(handle, data->stream_id); | |
4485 | ||
4486 | if (sample_type & PERF_SAMPLE_CPU) | |
4487 | perf_output_put(handle, data->cpu_entry); | |
4488 | ||
4489 | if (sample_type & PERF_SAMPLE_PERIOD) | |
4490 | perf_output_put(handle, data->period); | |
4491 | ||
4492 | if (sample_type & PERF_SAMPLE_READ) | |
cdd6c482 | 4493 | perf_output_read(handle, event); |
5622f295 MM |
4494 | |
4495 | if (sample_type & PERF_SAMPLE_CALLCHAIN) { | |
4496 | if (data->callchain) { | |
4497 | int size = 1; | |
4498 | ||
4499 | if (data->callchain) | |
4500 | size += data->callchain->nr; | |
4501 | ||
4502 | size *= sizeof(u64); | |
4503 | ||
76369139 | 4504 | __output_copy(handle, data->callchain, size); |
5622f295 MM |
4505 | } else { |
4506 | u64 nr = 0; | |
4507 | perf_output_put(handle, nr); | |
4508 | } | |
4509 | } | |
4510 | ||
4511 | if (sample_type & PERF_SAMPLE_RAW) { | |
4512 | if (data->raw) { | |
4513 | perf_output_put(handle, data->raw->size); | |
76369139 FW |
4514 | __output_copy(handle, data->raw->data, |
4515 | data->raw->size); | |
5622f295 MM |
4516 | } else { |
4517 | struct { | |
4518 | u32 size; | |
4519 | u32 data; | |
4520 | } raw = { | |
4521 | .size = sizeof(u32), | |
4522 | .data = 0, | |
4523 | }; | |
4524 | perf_output_put(handle, raw); | |
4525 | } | |
4526 | } | |
a7ac67ea | 4527 | |
bce38cd5 SE |
4528 | if (sample_type & PERF_SAMPLE_BRANCH_STACK) { |
4529 | if (data->br_stack) { | |
4530 | size_t size; | |
4531 | ||
4532 | size = data->br_stack->nr | |
4533 | * sizeof(struct perf_branch_entry); | |
4534 | ||
4535 | perf_output_put(handle, data->br_stack->nr); | |
4536 | perf_output_copy(handle, data->br_stack->entries, size); | |
4537 | } else { | |
4538 | /* | |
4539 | * we always store at least the value of nr | |
4540 | */ | |
4541 | u64 nr = 0; | |
4542 | perf_output_put(handle, nr); | |
4543 | } | |
4544 | } | |
4018994f JO |
4545 | |
4546 | if (sample_type & PERF_SAMPLE_REGS_USER) { | |
4547 | u64 abi = data->regs_user.abi; | |
4548 | ||
4549 | /* | |
4550 | * If there are no regs to dump, notice it through | |
4551 | * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE). | |
4552 | */ | |
4553 | perf_output_put(handle, abi); | |
4554 | ||
4555 | if (abi) { | |
4556 | u64 mask = event->attr.sample_regs_user; | |
4557 | perf_output_sample_regs(handle, | |
4558 | data->regs_user.regs, | |
4559 | mask); | |
4560 | } | |
4561 | } | |
c5ebcedb | 4562 | |
a5cdd40c | 4563 | if (sample_type & PERF_SAMPLE_STACK_USER) { |
c5ebcedb JO |
4564 | perf_output_sample_ustack(handle, |
4565 | data->stack_user_size, | |
4566 | data->regs_user.regs); | |
a5cdd40c | 4567 | } |
c3feedf2 AK |
4568 | |
4569 | if (sample_type & PERF_SAMPLE_WEIGHT) | |
4570 | perf_output_put(handle, data->weight); | |
d6be9ad6 SE |
4571 | |
4572 | if (sample_type & PERF_SAMPLE_DATA_SRC) | |
4573 | perf_output_put(handle, data->data_src.val); | |
a5cdd40c PZ |
4574 | |
4575 | if (!event->attr.watermark) { | |
4576 | int wakeup_events = event->attr.wakeup_events; | |
4577 | ||
4578 | if (wakeup_events) { | |
4579 | struct ring_buffer *rb = handle->rb; | |
4580 | int events = local_inc_return(&rb->events); | |
4581 | ||
4582 | if (events >= wakeup_events) { | |
4583 | local_sub(wakeup_events, &rb->events); | |
4584 | local_inc(&rb->wakeup); | |
4585 | } | |
4586 | } | |
4587 | } | |
5622f295 MM |
4588 | } |
4589 | ||
4590 | void perf_prepare_sample(struct perf_event_header *header, | |
4591 | struct perf_sample_data *data, | |
cdd6c482 | 4592 | struct perf_event *event, |
5622f295 | 4593 | struct pt_regs *regs) |
7b732a75 | 4594 | { |
cdd6c482 | 4595 | u64 sample_type = event->attr.sample_type; |
7b732a75 | 4596 | |
cdd6c482 | 4597 | header->type = PERF_RECORD_SAMPLE; |
c320c7b7 | 4598 | header->size = sizeof(*header) + event->header_size; |
5622f295 MM |
4599 | |
4600 | header->misc = 0; | |
4601 | header->misc |= perf_misc_flags(regs); | |
6fab0192 | 4602 | |
c980d109 | 4603 | __perf_event_header__init_id(header, data, event); |
6844c09d | 4604 | |
c320c7b7 | 4605 | if (sample_type & PERF_SAMPLE_IP) |
5622f295 MM |
4606 | data->ip = perf_instruction_pointer(regs); |
4607 | ||
b23f3325 | 4608 | if (sample_type & PERF_SAMPLE_CALLCHAIN) { |
5622f295 | 4609 | int size = 1; |
394ee076 | 4610 | |
e6dab5ff | 4611 | data->callchain = perf_callchain(event, regs); |
5622f295 MM |
4612 | |
4613 | if (data->callchain) | |
4614 | size += data->callchain->nr; | |
4615 | ||
4616 | header->size += size * sizeof(u64); | |
394ee076 PZ |
4617 | } |
4618 | ||
3a43ce68 | 4619 | if (sample_type & PERF_SAMPLE_RAW) { |
a044560c PZ |
4620 | int size = sizeof(u32); |
4621 | ||
4622 | if (data->raw) | |
4623 | size += data->raw->size; | |
4624 | else | |
4625 | size += sizeof(u32); | |
4626 | ||
4627 | WARN_ON_ONCE(size & (sizeof(u64)-1)); | |
5622f295 | 4628 | header->size += size; |
7f453c24 | 4629 | } |
bce38cd5 SE |
4630 | |
4631 | if (sample_type & PERF_SAMPLE_BRANCH_STACK) { | |
4632 | int size = sizeof(u64); /* nr */ | |
4633 | if (data->br_stack) { | |
4634 | size += data->br_stack->nr | |
4635 | * sizeof(struct perf_branch_entry); | |
4636 | } | |
4637 | header->size += size; | |
4638 | } | |
4018994f JO |
4639 | |
4640 | if (sample_type & PERF_SAMPLE_REGS_USER) { | |
4641 | /* regs dump ABI info */ | |
4642 | int size = sizeof(u64); | |
4643 | ||
4644 | perf_sample_regs_user(&data->regs_user, regs); | |
4645 | ||
4646 | if (data->regs_user.regs) { | |
4647 | u64 mask = event->attr.sample_regs_user; | |
4648 | size += hweight64(mask) * sizeof(u64); | |
4649 | } | |
4650 | ||
4651 | header->size += size; | |
4652 | } | |
c5ebcedb JO |
4653 | |
4654 | if (sample_type & PERF_SAMPLE_STACK_USER) { | |
4655 | /* | |
4656 | * Either we need PERF_SAMPLE_STACK_USER bit to be allways | |
4657 | * processed as the last one or have additional check added | |
4658 | * in case new sample type is added, because we could eat | |
4659 | * up the rest of the sample size. | |
4660 | */ | |
4661 | struct perf_regs_user *uregs = &data->regs_user; | |
4662 | u16 stack_size = event->attr.sample_stack_user; | |
4663 | u16 size = sizeof(u64); | |
4664 | ||
4665 | if (!uregs->abi) | |
4666 | perf_sample_regs_user(uregs, regs); | |
4667 | ||
4668 | stack_size = perf_sample_ustack_size(stack_size, header->size, | |
4669 | uregs->regs); | |
4670 | ||
4671 | /* | |
4672 | * If there is something to dump, add space for the dump | |
4673 | * itself and for the field that tells the dynamic size, | |
4674 | * which is how many have been actually dumped. | |
4675 | */ | |
4676 | if (stack_size) | |
4677 | size += sizeof(u64) + stack_size; | |
4678 | ||
4679 | data->stack_user_size = stack_size; | |
4680 | header->size += size; | |
4681 | } | |
5622f295 | 4682 | } |
7f453c24 | 4683 | |
a8b0ca17 | 4684 | static void perf_event_output(struct perf_event *event, |
5622f295 MM |
4685 | struct perf_sample_data *data, |
4686 | struct pt_regs *regs) | |
4687 | { | |
4688 | struct perf_output_handle handle; | |
4689 | struct perf_event_header header; | |
689802b2 | 4690 | |
927c7a9e FW |
4691 | /* protect the callchain buffers */ |
4692 | rcu_read_lock(); | |
4693 | ||
cdd6c482 | 4694 | perf_prepare_sample(&header, data, event, regs); |
5c148194 | 4695 | |
a7ac67ea | 4696 | if (perf_output_begin(&handle, event, header.size)) |
927c7a9e | 4697 | goto exit; |
0322cd6e | 4698 | |
cdd6c482 | 4699 | perf_output_sample(&handle, &header, data, event); |
f413cdb8 | 4700 | |
8a057d84 | 4701 | perf_output_end(&handle); |
927c7a9e FW |
4702 | |
4703 | exit: | |
4704 | rcu_read_unlock(); | |
0322cd6e PZ |
4705 | } |
4706 | ||
38b200d6 | 4707 | /* |
cdd6c482 | 4708 | * read event_id |
38b200d6 PZ |
4709 | */ |
4710 | ||
4711 | struct perf_read_event { | |
4712 | struct perf_event_header header; | |
4713 | ||
4714 | u32 pid; | |
4715 | u32 tid; | |
38b200d6 PZ |
4716 | }; |
4717 | ||
4718 | static void | |
cdd6c482 | 4719 | perf_event_read_event(struct perf_event *event, |
38b200d6 PZ |
4720 | struct task_struct *task) |
4721 | { | |
4722 | struct perf_output_handle handle; | |
c980d109 | 4723 | struct perf_sample_data sample; |
dfc65094 | 4724 | struct perf_read_event read_event = { |
38b200d6 | 4725 | .header = { |
cdd6c482 | 4726 | .type = PERF_RECORD_READ, |
38b200d6 | 4727 | .misc = 0, |
c320c7b7 | 4728 | .size = sizeof(read_event) + event->read_size, |
38b200d6 | 4729 | }, |
cdd6c482 IM |
4730 | .pid = perf_event_pid(event, task), |
4731 | .tid = perf_event_tid(event, task), | |
38b200d6 | 4732 | }; |
3dab77fb | 4733 | int ret; |
38b200d6 | 4734 | |
c980d109 | 4735 | perf_event_header__init_id(&read_event.header, &sample, event); |
a7ac67ea | 4736 | ret = perf_output_begin(&handle, event, read_event.header.size); |
38b200d6 PZ |
4737 | if (ret) |
4738 | return; | |
4739 | ||
dfc65094 | 4740 | perf_output_put(&handle, read_event); |
cdd6c482 | 4741 | perf_output_read(&handle, event); |
c980d109 | 4742 | perf_event__output_id_sample(event, &handle, &sample); |
3dab77fb | 4743 | |
38b200d6 PZ |
4744 | perf_output_end(&handle); |
4745 | } | |
4746 | ||
52d857a8 JO |
4747 | typedef void (perf_event_aux_output_cb)(struct perf_event *event, void *data); |
4748 | ||
4749 | static void | |
4750 | perf_event_aux_ctx(struct perf_event_context *ctx, | |
52d857a8 JO |
4751 | perf_event_aux_output_cb output, |
4752 | void *data) | |
4753 | { | |
4754 | struct perf_event *event; | |
4755 | ||
4756 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { | |
4757 | if (event->state < PERF_EVENT_STATE_INACTIVE) | |
4758 | continue; | |
4759 | if (!event_filter_match(event)) | |
4760 | continue; | |
67516844 | 4761 | output(event, data); |
52d857a8 JO |
4762 | } |
4763 | } | |
4764 | ||
4765 | static void | |
67516844 | 4766 | perf_event_aux(perf_event_aux_output_cb output, void *data, |
52d857a8 JO |
4767 | struct perf_event_context *task_ctx) |
4768 | { | |
4769 | struct perf_cpu_context *cpuctx; | |
4770 | struct perf_event_context *ctx; | |
4771 | struct pmu *pmu; | |
4772 | int ctxn; | |
4773 | ||
4774 | rcu_read_lock(); | |
4775 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
4776 | cpuctx = get_cpu_ptr(pmu->pmu_cpu_context); | |
4777 | if (cpuctx->unique_pmu != pmu) | |
4778 | goto next; | |
67516844 | 4779 | perf_event_aux_ctx(&cpuctx->ctx, output, data); |
52d857a8 JO |
4780 | if (task_ctx) |
4781 | goto next; | |
4782 | ctxn = pmu->task_ctx_nr; | |
4783 | if (ctxn < 0) | |
4784 | goto next; | |
4785 | ctx = rcu_dereference(current->perf_event_ctxp[ctxn]); | |
4786 | if (ctx) | |
67516844 | 4787 | perf_event_aux_ctx(ctx, output, data); |
52d857a8 JO |
4788 | next: |
4789 | put_cpu_ptr(pmu->pmu_cpu_context); | |
4790 | } | |
4791 | ||
4792 | if (task_ctx) { | |
4793 | preempt_disable(); | |
67516844 | 4794 | perf_event_aux_ctx(task_ctx, output, data); |
52d857a8 JO |
4795 | preempt_enable(); |
4796 | } | |
4797 | rcu_read_unlock(); | |
4798 | } | |
4799 | ||
60313ebe | 4800 | /* |
9f498cc5 PZ |
4801 | * task tracking -- fork/exit |
4802 | * | |
13d7a241 | 4803 | * enabled by: attr.comm | attr.mmap | attr.mmap2 | attr.mmap_data | attr.task |
60313ebe PZ |
4804 | */ |
4805 | ||
9f498cc5 | 4806 | struct perf_task_event { |
3a80b4a3 | 4807 | struct task_struct *task; |
cdd6c482 | 4808 | struct perf_event_context *task_ctx; |
60313ebe PZ |
4809 | |
4810 | struct { | |
4811 | struct perf_event_header header; | |
4812 | ||
4813 | u32 pid; | |
4814 | u32 ppid; | |
9f498cc5 PZ |
4815 | u32 tid; |
4816 | u32 ptid; | |
393b2ad8 | 4817 | u64 time; |
cdd6c482 | 4818 | } event_id; |
60313ebe PZ |
4819 | }; |
4820 | ||
67516844 JO |
4821 | static int perf_event_task_match(struct perf_event *event) |
4822 | { | |
13d7a241 SE |
4823 | return event->attr.comm || event->attr.mmap || |
4824 | event->attr.mmap2 || event->attr.mmap_data || | |
4825 | event->attr.task; | |
67516844 JO |
4826 | } |
4827 | ||
cdd6c482 | 4828 | static void perf_event_task_output(struct perf_event *event, |
52d857a8 | 4829 | void *data) |
60313ebe | 4830 | { |
52d857a8 | 4831 | struct perf_task_event *task_event = data; |
60313ebe | 4832 | struct perf_output_handle handle; |
c980d109 | 4833 | struct perf_sample_data sample; |
9f498cc5 | 4834 | struct task_struct *task = task_event->task; |
c980d109 | 4835 | int ret, size = task_event->event_id.header.size; |
8bb39f9a | 4836 | |
67516844 JO |
4837 | if (!perf_event_task_match(event)) |
4838 | return; | |
4839 | ||
c980d109 | 4840 | perf_event_header__init_id(&task_event->event_id.header, &sample, event); |
60313ebe | 4841 | |
c980d109 | 4842 | ret = perf_output_begin(&handle, event, |
a7ac67ea | 4843 | task_event->event_id.header.size); |
ef60777c | 4844 | if (ret) |
c980d109 | 4845 | goto out; |
60313ebe | 4846 | |
cdd6c482 IM |
4847 | task_event->event_id.pid = perf_event_pid(event, task); |
4848 | task_event->event_id.ppid = perf_event_pid(event, current); | |
60313ebe | 4849 | |
cdd6c482 IM |
4850 | task_event->event_id.tid = perf_event_tid(event, task); |
4851 | task_event->event_id.ptid = perf_event_tid(event, current); | |
9f498cc5 | 4852 | |
cdd6c482 | 4853 | perf_output_put(&handle, task_event->event_id); |
393b2ad8 | 4854 | |
c980d109 ACM |
4855 | perf_event__output_id_sample(event, &handle, &sample); |
4856 | ||
60313ebe | 4857 | perf_output_end(&handle); |
c980d109 ACM |
4858 | out: |
4859 | task_event->event_id.header.size = size; | |
60313ebe PZ |
4860 | } |
4861 | ||
cdd6c482 IM |
4862 | static void perf_event_task(struct task_struct *task, |
4863 | struct perf_event_context *task_ctx, | |
3a80b4a3 | 4864 | int new) |
60313ebe | 4865 | { |
9f498cc5 | 4866 | struct perf_task_event task_event; |
60313ebe | 4867 | |
cdd6c482 IM |
4868 | if (!atomic_read(&nr_comm_events) && |
4869 | !atomic_read(&nr_mmap_events) && | |
4870 | !atomic_read(&nr_task_events)) | |
60313ebe PZ |
4871 | return; |
4872 | ||
9f498cc5 | 4873 | task_event = (struct perf_task_event){ |
3a80b4a3 PZ |
4874 | .task = task, |
4875 | .task_ctx = task_ctx, | |
cdd6c482 | 4876 | .event_id = { |
60313ebe | 4877 | .header = { |
cdd6c482 | 4878 | .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT, |
573402db | 4879 | .misc = 0, |
cdd6c482 | 4880 | .size = sizeof(task_event.event_id), |
60313ebe | 4881 | }, |
573402db PZ |
4882 | /* .pid */ |
4883 | /* .ppid */ | |
9f498cc5 PZ |
4884 | /* .tid */ |
4885 | /* .ptid */ | |
6f93d0a7 | 4886 | .time = perf_clock(), |
60313ebe PZ |
4887 | }, |
4888 | }; | |
4889 | ||
67516844 | 4890 | perf_event_aux(perf_event_task_output, |
52d857a8 JO |
4891 | &task_event, |
4892 | task_ctx); | |
9f498cc5 PZ |
4893 | } |
4894 | ||
cdd6c482 | 4895 | void perf_event_fork(struct task_struct *task) |
9f498cc5 | 4896 | { |
cdd6c482 | 4897 | perf_event_task(task, NULL, 1); |
60313ebe PZ |
4898 | } |
4899 | ||
8d1b2d93 PZ |
4900 | /* |
4901 | * comm tracking | |
4902 | */ | |
4903 | ||
4904 | struct perf_comm_event { | |
22a4f650 IM |
4905 | struct task_struct *task; |
4906 | char *comm; | |
8d1b2d93 PZ |
4907 | int comm_size; |
4908 | ||
4909 | struct { | |
4910 | struct perf_event_header header; | |
4911 | ||
4912 | u32 pid; | |
4913 | u32 tid; | |
cdd6c482 | 4914 | } event_id; |
8d1b2d93 PZ |
4915 | }; |
4916 | ||
67516844 JO |
4917 | static int perf_event_comm_match(struct perf_event *event) |
4918 | { | |
4919 | return event->attr.comm; | |
4920 | } | |
4921 | ||
cdd6c482 | 4922 | static void perf_event_comm_output(struct perf_event *event, |
52d857a8 | 4923 | void *data) |
8d1b2d93 | 4924 | { |
52d857a8 | 4925 | struct perf_comm_event *comm_event = data; |
8d1b2d93 | 4926 | struct perf_output_handle handle; |
c980d109 | 4927 | struct perf_sample_data sample; |
cdd6c482 | 4928 | int size = comm_event->event_id.header.size; |
c980d109 ACM |
4929 | int ret; |
4930 | ||
67516844 JO |
4931 | if (!perf_event_comm_match(event)) |
4932 | return; | |
4933 | ||
c980d109 ACM |
4934 | perf_event_header__init_id(&comm_event->event_id.header, &sample, event); |
4935 | ret = perf_output_begin(&handle, event, | |
a7ac67ea | 4936 | comm_event->event_id.header.size); |
8d1b2d93 PZ |
4937 | |
4938 | if (ret) | |
c980d109 | 4939 | goto out; |
8d1b2d93 | 4940 | |
cdd6c482 IM |
4941 | comm_event->event_id.pid = perf_event_pid(event, comm_event->task); |
4942 | comm_event->event_id.tid = perf_event_tid(event, comm_event->task); | |
709e50cf | 4943 | |
cdd6c482 | 4944 | perf_output_put(&handle, comm_event->event_id); |
76369139 | 4945 | __output_copy(&handle, comm_event->comm, |
8d1b2d93 | 4946 | comm_event->comm_size); |
c980d109 ACM |
4947 | |
4948 | perf_event__output_id_sample(event, &handle, &sample); | |
4949 | ||
8d1b2d93 | 4950 | perf_output_end(&handle); |
c980d109 ACM |
4951 | out: |
4952 | comm_event->event_id.header.size = size; | |
8d1b2d93 PZ |
4953 | } |
4954 | ||
cdd6c482 | 4955 | static void perf_event_comm_event(struct perf_comm_event *comm_event) |
8d1b2d93 | 4956 | { |
413ee3b4 | 4957 | char comm[TASK_COMM_LEN]; |
8d1b2d93 | 4958 | unsigned int size; |
8d1b2d93 | 4959 | |
413ee3b4 | 4960 | memset(comm, 0, sizeof(comm)); |
96b02d78 | 4961 | strlcpy(comm, comm_event->task->comm, sizeof(comm)); |
888fcee0 | 4962 | size = ALIGN(strlen(comm)+1, sizeof(u64)); |
8d1b2d93 PZ |
4963 | |
4964 | comm_event->comm = comm; | |
4965 | comm_event->comm_size = size; | |
4966 | ||
cdd6c482 | 4967 | comm_event->event_id.header.size = sizeof(comm_event->event_id) + size; |
8dc85d54 | 4968 | |
67516844 | 4969 | perf_event_aux(perf_event_comm_output, |
52d857a8 JO |
4970 | comm_event, |
4971 | NULL); | |
8d1b2d93 PZ |
4972 | } |
4973 | ||
cdd6c482 | 4974 | void perf_event_comm(struct task_struct *task) |
8d1b2d93 | 4975 | { |
9ee318a7 | 4976 | struct perf_comm_event comm_event; |
8dc85d54 PZ |
4977 | struct perf_event_context *ctx; |
4978 | int ctxn; | |
9ee318a7 | 4979 | |
c79aa0d9 | 4980 | rcu_read_lock(); |
8dc85d54 PZ |
4981 | for_each_task_context_nr(ctxn) { |
4982 | ctx = task->perf_event_ctxp[ctxn]; | |
4983 | if (!ctx) | |
4984 | continue; | |
9ee318a7 | 4985 | |
8dc85d54 PZ |
4986 | perf_event_enable_on_exec(ctx); |
4987 | } | |
c79aa0d9 | 4988 | rcu_read_unlock(); |
9ee318a7 | 4989 | |
cdd6c482 | 4990 | if (!atomic_read(&nr_comm_events)) |
9ee318a7 | 4991 | return; |
a63eaf34 | 4992 | |
9ee318a7 | 4993 | comm_event = (struct perf_comm_event){ |
8d1b2d93 | 4994 | .task = task, |
573402db PZ |
4995 | /* .comm */ |
4996 | /* .comm_size */ | |
cdd6c482 | 4997 | .event_id = { |
573402db | 4998 | .header = { |
cdd6c482 | 4999 | .type = PERF_RECORD_COMM, |
573402db PZ |
5000 | .misc = 0, |
5001 | /* .size */ | |
5002 | }, | |
5003 | /* .pid */ | |
5004 | /* .tid */ | |
8d1b2d93 PZ |
5005 | }, |
5006 | }; | |
5007 | ||
cdd6c482 | 5008 | perf_event_comm_event(&comm_event); |
8d1b2d93 PZ |
5009 | } |
5010 | ||
0a4a9391 PZ |
5011 | /* |
5012 | * mmap tracking | |
5013 | */ | |
5014 | ||
5015 | struct perf_mmap_event { | |
089dd79d PZ |
5016 | struct vm_area_struct *vma; |
5017 | ||
5018 | const char *file_name; | |
5019 | int file_size; | |
13d7a241 SE |
5020 | int maj, min; |
5021 | u64 ino; | |
5022 | u64 ino_generation; | |
0a4a9391 PZ |
5023 | |
5024 | struct { | |
5025 | struct perf_event_header header; | |
5026 | ||
5027 | u32 pid; | |
5028 | u32 tid; | |
5029 | u64 start; | |
5030 | u64 len; | |
5031 | u64 pgoff; | |
cdd6c482 | 5032 | } event_id; |
0a4a9391 PZ |
5033 | }; |
5034 | ||
67516844 JO |
5035 | static int perf_event_mmap_match(struct perf_event *event, |
5036 | void *data) | |
5037 | { | |
5038 | struct perf_mmap_event *mmap_event = data; | |
5039 | struct vm_area_struct *vma = mmap_event->vma; | |
5040 | int executable = vma->vm_flags & VM_EXEC; | |
5041 | ||
5042 | return (!executable && event->attr.mmap_data) || | |
13d7a241 | 5043 | (executable && (event->attr.mmap || event->attr.mmap2)); |
67516844 JO |
5044 | } |
5045 | ||
cdd6c482 | 5046 | static void perf_event_mmap_output(struct perf_event *event, |
52d857a8 | 5047 | void *data) |
0a4a9391 | 5048 | { |
52d857a8 | 5049 | struct perf_mmap_event *mmap_event = data; |
0a4a9391 | 5050 | struct perf_output_handle handle; |
c980d109 | 5051 | struct perf_sample_data sample; |
cdd6c482 | 5052 | int size = mmap_event->event_id.header.size; |
c980d109 | 5053 | int ret; |
0a4a9391 | 5054 | |
67516844 JO |
5055 | if (!perf_event_mmap_match(event, data)) |
5056 | return; | |
5057 | ||
13d7a241 SE |
5058 | if (event->attr.mmap2) { |
5059 | mmap_event->event_id.header.type = PERF_RECORD_MMAP2; | |
5060 | mmap_event->event_id.header.size += sizeof(mmap_event->maj); | |
5061 | mmap_event->event_id.header.size += sizeof(mmap_event->min); | |
5062 | mmap_event->event_id.header.size += sizeof(mmap_event->ino); | |
d008d525 | 5063 | mmap_event->event_id.header.size += sizeof(mmap_event->ino_generation); |
13d7a241 SE |
5064 | } |
5065 | ||
c980d109 ACM |
5066 | perf_event_header__init_id(&mmap_event->event_id.header, &sample, event); |
5067 | ret = perf_output_begin(&handle, event, | |
a7ac67ea | 5068 | mmap_event->event_id.header.size); |
0a4a9391 | 5069 | if (ret) |
c980d109 | 5070 | goto out; |
0a4a9391 | 5071 | |
cdd6c482 IM |
5072 | mmap_event->event_id.pid = perf_event_pid(event, current); |
5073 | mmap_event->event_id.tid = perf_event_tid(event, current); | |
709e50cf | 5074 | |
cdd6c482 | 5075 | perf_output_put(&handle, mmap_event->event_id); |
13d7a241 SE |
5076 | |
5077 | if (event->attr.mmap2) { | |
5078 | perf_output_put(&handle, mmap_event->maj); | |
5079 | perf_output_put(&handle, mmap_event->min); | |
5080 | perf_output_put(&handle, mmap_event->ino); | |
5081 | perf_output_put(&handle, mmap_event->ino_generation); | |
5082 | } | |
5083 | ||
76369139 | 5084 | __output_copy(&handle, mmap_event->file_name, |
0a4a9391 | 5085 | mmap_event->file_size); |
c980d109 ACM |
5086 | |
5087 | perf_event__output_id_sample(event, &handle, &sample); | |
5088 | ||
78d613eb | 5089 | perf_output_end(&handle); |
c980d109 ACM |
5090 | out: |
5091 | mmap_event->event_id.header.size = size; | |
0a4a9391 PZ |
5092 | } |
5093 | ||
cdd6c482 | 5094 | static void perf_event_mmap_event(struct perf_mmap_event *mmap_event) |
0a4a9391 | 5095 | { |
089dd79d PZ |
5096 | struct vm_area_struct *vma = mmap_event->vma; |
5097 | struct file *file = vma->vm_file; | |
13d7a241 SE |
5098 | int maj = 0, min = 0; |
5099 | u64 ino = 0, gen = 0; | |
0a4a9391 PZ |
5100 | unsigned int size; |
5101 | char tmp[16]; | |
5102 | char *buf = NULL; | |
089dd79d | 5103 | const char *name; |
0a4a9391 | 5104 | |
413ee3b4 AB |
5105 | memset(tmp, 0, sizeof(tmp)); |
5106 | ||
0a4a9391 | 5107 | if (file) { |
13d7a241 SE |
5108 | struct inode *inode; |
5109 | dev_t dev; | |
413ee3b4 | 5110 | /* |
76369139 | 5111 | * d_path works from the end of the rb backwards, so we |
413ee3b4 AB |
5112 | * need to add enough zero bytes after the string to handle |
5113 | * the 64bit alignment we do later. | |
5114 | */ | |
5115 | buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL); | |
0a4a9391 PZ |
5116 | if (!buf) { |
5117 | name = strncpy(tmp, "//enomem", sizeof(tmp)); | |
5118 | goto got_name; | |
5119 | } | |
d3d21c41 | 5120 | name = d_path(&file->f_path, buf, PATH_MAX); |
0a4a9391 PZ |
5121 | if (IS_ERR(name)) { |
5122 | name = strncpy(tmp, "//toolong", sizeof(tmp)); | |
5123 | goto got_name; | |
5124 | } | |
13d7a241 SE |
5125 | inode = file_inode(vma->vm_file); |
5126 | dev = inode->i_sb->s_dev; | |
5127 | ino = inode->i_ino; | |
5128 | gen = inode->i_generation; | |
5129 | maj = MAJOR(dev); | |
5130 | min = MINOR(dev); | |
5131 | ||
0a4a9391 | 5132 | } else { |
413ee3b4 AB |
5133 | if (arch_vma_name(mmap_event->vma)) { |
5134 | name = strncpy(tmp, arch_vma_name(mmap_event->vma), | |
c97847d2 CG |
5135 | sizeof(tmp) - 1); |
5136 | tmp[sizeof(tmp) - 1] = '\0'; | |
089dd79d | 5137 | goto got_name; |
413ee3b4 | 5138 | } |
089dd79d PZ |
5139 | |
5140 | if (!vma->vm_mm) { | |
5141 | name = strncpy(tmp, "[vdso]", sizeof(tmp)); | |
5142 | goto got_name; | |
3af9e859 EM |
5143 | } else if (vma->vm_start <= vma->vm_mm->start_brk && |
5144 | vma->vm_end >= vma->vm_mm->brk) { | |
5145 | name = strncpy(tmp, "[heap]", sizeof(tmp)); | |
5146 | goto got_name; | |
5147 | } else if (vma->vm_start <= vma->vm_mm->start_stack && | |
5148 | vma->vm_end >= vma->vm_mm->start_stack) { | |
5149 | name = strncpy(tmp, "[stack]", sizeof(tmp)); | |
5150 | goto got_name; | |
089dd79d PZ |
5151 | } |
5152 | ||
0a4a9391 PZ |
5153 | name = strncpy(tmp, "//anon", sizeof(tmp)); |
5154 | goto got_name; | |
5155 | } | |
5156 | ||
5157 | got_name: | |
888fcee0 | 5158 | size = ALIGN(strlen(name)+1, sizeof(u64)); |
0a4a9391 PZ |
5159 | |
5160 | mmap_event->file_name = name; | |
5161 | mmap_event->file_size = size; | |
13d7a241 SE |
5162 | mmap_event->maj = maj; |
5163 | mmap_event->min = min; | |
5164 | mmap_event->ino = ino; | |
5165 | mmap_event->ino_generation = gen; | |
0a4a9391 | 5166 | |
2fe85427 SE |
5167 | if (!(vma->vm_flags & VM_EXEC)) |
5168 | mmap_event->event_id.header.misc |= PERF_RECORD_MISC_MMAP_DATA; | |
5169 | ||
cdd6c482 | 5170 | mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size; |
0a4a9391 | 5171 | |
67516844 | 5172 | perf_event_aux(perf_event_mmap_output, |
52d857a8 JO |
5173 | mmap_event, |
5174 | NULL); | |
665c2142 | 5175 | |
0a4a9391 PZ |
5176 | kfree(buf); |
5177 | } | |
5178 | ||
3af9e859 | 5179 | void perf_event_mmap(struct vm_area_struct *vma) |
0a4a9391 | 5180 | { |
9ee318a7 PZ |
5181 | struct perf_mmap_event mmap_event; |
5182 | ||
cdd6c482 | 5183 | if (!atomic_read(&nr_mmap_events)) |
9ee318a7 PZ |
5184 | return; |
5185 | ||
5186 | mmap_event = (struct perf_mmap_event){ | |
089dd79d | 5187 | .vma = vma, |
573402db PZ |
5188 | /* .file_name */ |
5189 | /* .file_size */ | |
cdd6c482 | 5190 | .event_id = { |
573402db | 5191 | .header = { |
cdd6c482 | 5192 | .type = PERF_RECORD_MMAP, |
39447b38 | 5193 | .misc = PERF_RECORD_MISC_USER, |
573402db PZ |
5194 | /* .size */ |
5195 | }, | |
5196 | /* .pid */ | |
5197 | /* .tid */ | |
089dd79d PZ |
5198 | .start = vma->vm_start, |
5199 | .len = vma->vm_end - vma->vm_start, | |
3a0304e9 | 5200 | .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT, |
0a4a9391 | 5201 | }, |
13d7a241 SE |
5202 | /* .maj (attr_mmap2 only) */ |
5203 | /* .min (attr_mmap2 only) */ | |
5204 | /* .ino (attr_mmap2 only) */ | |
5205 | /* .ino_generation (attr_mmap2 only) */ | |
0a4a9391 PZ |
5206 | }; |
5207 | ||
cdd6c482 | 5208 | perf_event_mmap_event(&mmap_event); |
0a4a9391 PZ |
5209 | } |
5210 | ||
a78ac325 PZ |
5211 | /* |
5212 | * IRQ throttle logging | |
5213 | */ | |
5214 | ||
cdd6c482 | 5215 | static void perf_log_throttle(struct perf_event *event, int enable) |
a78ac325 PZ |
5216 | { |
5217 | struct perf_output_handle handle; | |
c980d109 | 5218 | struct perf_sample_data sample; |
a78ac325 PZ |
5219 | int ret; |
5220 | ||
5221 | struct { | |
5222 | struct perf_event_header header; | |
5223 | u64 time; | |
cca3f454 | 5224 | u64 id; |
7f453c24 | 5225 | u64 stream_id; |
a78ac325 PZ |
5226 | } throttle_event = { |
5227 | .header = { | |
cdd6c482 | 5228 | .type = PERF_RECORD_THROTTLE, |
a78ac325 PZ |
5229 | .misc = 0, |
5230 | .size = sizeof(throttle_event), | |
5231 | }, | |
def0a9b2 | 5232 | .time = perf_clock(), |
cdd6c482 IM |
5233 | .id = primary_event_id(event), |
5234 | .stream_id = event->id, | |
a78ac325 PZ |
5235 | }; |
5236 | ||
966ee4d6 | 5237 | if (enable) |
cdd6c482 | 5238 | throttle_event.header.type = PERF_RECORD_UNTHROTTLE; |
966ee4d6 | 5239 | |
c980d109 ACM |
5240 | perf_event_header__init_id(&throttle_event.header, &sample, event); |
5241 | ||
5242 | ret = perf_output_begin(&handle, event, | |
a7ac67ea | 5243 | throttle_event.header.size); |
a78ac325 PZ |
5244 | if (ret) |
5245 | return; | |
5246 | ||
5247 | perf_output_put(&handle, throttle_event); | |
c980d109 | 5248 | perf_event__output_id_sample(event, &handle, &sample); |
a78ac325 PZ |
5249 | perf_output_end(&handle); |
5250 | } | |
5251 | ||
f6c7d5fe | 5252 | /* |
cdd6c482 | 5253 | * Generic event overflow handling, sampling. |
f6c7d5fe PZ |
5254 | */ |
5255 | ||
a8b0ca17 | 5256 | static int __perf_event_overflow(struct perf_event *event, |
5622f295 MM |
5257 | int throttle, struct perf_sample_data *data, |
5258 | struct pt_regs *regs) | |
f6c7d5fe | 5259 | { |
cdd6c482 IM |
5260 | int events = atomic_read(&event->event_limit); |
5261 | struct hw_perf_event *hwc = &event->hw; | |
e050e3f0 | 5262 | u64 seq; |
79f14641 PZ |
5263 | int ret = 0; |
5264 | ||
96398826 PZ |
5265 | /* |
5266 | * Non-sampling counters might still use the PMI to fold short | |
5267 | * hardware counters, ignore those. | |
5268 | */ | |
5269 | if (unlikely(!is_sampling_event(event))) | |
5270 | return 0; | |
5271 | ||
e050e3f0 SE |
5272 | seq = __this_cpu_read(perf_throttled_seq); |
5273 | if (seq != hwc->interrupts_seq) { | |
5274 | hwc->interrupts_seq = seq; | |
5275 | hwc->interrupts = 1; | |
5276 | } else { | |
5277 | hwc->interrupts++; | |
5278 | if (unlikely(throttle | |
5279 | && hwc->interrupts >= max_samples_per_tick)) { | |
5280 | __this_cpu_inc(perf_throttled_count); | |
163ec435 PZ |
5281 | hwc->interrupts = MAX_INTERRUPTS; |
5282 | perf_log_throttle(event, 0); | |
d84153d6 | 5283 | tick_nohz_full_kick(); |
a78ac325 PZ |
5284 | ret = 1; |
5285 | } | |
e050e3f0 | 5286 | } |
60db5e09 | 5287 | |
cdd6c482 | 5288 | if (event->attr.freq) { |
def0a9b2 | 5289 | u64 now = perf_clock(); |
abd50713 | 5290 | s64 delta = now - hwc->freq_time_stamp; |
bd2b5b12 | 5291 | |
abd50713 | 5292 | hwc->freq_time_stamp = now; |
bd2b5b12 | 5293 | |
abd50713 | 5294 | if (delta > 0 && delta < 2*TICK_NSEC) |
f39d47ff | 5295 | perf_adjust_period(event, delta, hwc->last_period, true); |
bd2b5b12 PZ |
5296 | } |
5297 | ||
2023b359 PZ |
5298 | /* |
5299 | * XXX event_limit might not quite work as expected on inherited | |
cdd6c482 | 5300 | * events |
2023b359 PZ |
5301 | */ |
5302 | ||
cdd6c482 IM |
5303 | event->pending_kill = POLL_IN; |
5304 | if (events && atomic_dec_and_test(&event->event_limit)) { | |
79f14641 | 5305 | ret = 1; |
cdd6c482 | 5306 | event->pending_kill = POLL_HUP; |
a8b0ca17 PZ |
5307 | event->pending_disable = 1; |
5308 | irq_work_queue(&event->pending); | |
79f14641 PZ |
5309 | } |
5310 | ||
453f19ee | 5311 | if (event->overflow_handler) |
a8b0ca17 | 5312 | event->overflow_handler(event, data, regs); |
453f19ee | 5313 | else |
a8b0ca17 | 5314 | perf_event_output(event, data, regs); |
453f19ee | 5315 | |
f506b3dc | 5316 | if (event->fasync && event->pending_kill) { |
a8b0ca17 PZ |
5317 | event->pending_wakeup = 1; |
5318 | irq_work_queue(&event->pending); | |
f506b3dc PZ |
5319 | } |
5320 | ||
79f14641 | 5321 | return ret; |
f6c7d5fe PZ |
5322 | } |
5323 | ||
a8b0ca17 | 5324 | int perf_event_overflow(struct perf_event *event, |
5622f295 MM |
5325 | struct perf_sample_data *data, |
5326 | struct pt_regs *regs) | |
850bc73f | 5327 | { |
a8b0ca17 | 5328 | return __perf_event_overflow(event, 1, data, regs); |
850bc73f PZ |
5329 | } |
5330 | ||
15dbf27c | 5331 | /* |
cdd6c482 | 5332 | * Generic software event infrastructure |
15dbf27c PZ |
5333 | */ |
5334 | ||
b28ab83c PZ |
5335 | struct swevent_htable { |
5336 | struct swevent_hlist *swevent_hlist; | |
5337 | struct mutex hlist_mutex; | |
5338 | int hlist_refcount; | |
5339 | ||
5340 | /* Recursion avoidance in each contexts */ | |
5341 | int recursion[PERF_NR_CONTEXTS]; | |
5342 | }; | |
5343 | ||
5344 | static DEFINE_PER_CPU(struct swevent_htable, swevent_htable); | |
5345 | ||
7b4b6658 | 5346 | /* |
cdd6c482 IM |
5347 | * We directly increment event->count and keep a second value in |
5348 | * event->hw.period_left to count intervals. This period event | |
7b4b6658 PZ |
5349 | * is kept in the range [-sample_period, 0] so that we can use the |
5350 | * sign as trigger. | |
5351 | */ | |
5352 | ||
ab573844 | 5353 | u64 perf_swevent_set_period(struct perf_event *event) |
15dbf27c | 5354 | { |
cdd6c482 | 5355 | struct hw_perf_event *hwc = &event->hw; |
7b4b6658 PZ |
5356 | u64 period = hwc->last_period; |
5357 | u64 nr, offset; | |
5358 | s64 old, val; | |
5359 | ||
5360 | hwc->last_period = hwc->sample_period; | |
15dbf27c PZ |
5361 | |
5362 | again: | |
e7850595 | 5363 | old = val = local64_read(&hwc->period_left); |
7b4b6658 PZ |
5364 | if (val < 0) |
5365 | return 0; | |
15dbf27c | 5366 | |
7b4b6658 PZ |
5367 | nr = div64_u64(period + val, period); |
5368 | offset = nr * period; | |
5369 | val -= offset; | |
e7850595 | 5370 | if (local64_cmpxchg(&hwc->period_left, old, val) != old) |
7b4b6658 | 5371 | goto again; |
15dbf27c | 5372 | |
7b4b6658 | 5373 | return nr; |
15dbf27c PZ |
5374 | } |
5375 | ||
0cff784a | 5376 | static void perf_swevent_overflow(struct perf_event *event, u64 overflow, |
a8b0ca17 | 5377 | struct perf_sample_data *data, |
5622f295 | 5378 | struct pt_regs *regs) |
15dbf27c | 5379 | { |
cdd6c482 | 5380 | struct hw_perf_event *hwc = &event->hw; |
850bc73f | 5381 | int throttle = 0; |
15dbf27c | 5382 | |
0cff784a PZ |
5383 | if (!overflow) |
5384 | overflow = perf_swevent_set_period(event); | |
15dbf27c | 5385 | |
7b4b6658 PZ |
5386 | if (hwc->interrupts == MAX_INTERRUPTS) |
5387 | return; | |
15dbf27c | 5388 | |
7b4b6658 | 5389 | for (; overflow; overflow--) { |
a8b0ca17 | 5390 | if (__perf_event_overflow(event, throttle, |
5622f295 | 5391 | data, regs)) { |
7b4b6658 PZ |
5392 | /* |
5393 | * We inhibit the overflow from happening when | |
5394 | * hwc->interrupts == MAX_INTERRUPTS. | |
5395 | */ | |
5396 | break; | |
5397 | } | |
cf450a73 | 5398 | throttle = 1; |
7b4b6658 | 5399 | } |
15dbf27c PZ |
5400 | } |
5401 | ||
a4eaf7f1 | 5402 | static void perf_swevent_event(struct perf_event *event, u64 nr, |
a8b0ca17 | 5403 | struct perf_sample_data *data, |
5622f295 | 5404 | struct pt_regs *regs) |
7b4b6658 | 5405 | { |
cdd6c482 | 5406 | struct hw_perf_event *hwc = &event->hw; |
d6d020e9 | 5407 | |
e7850595 | 5408 | local64_add(nr, &event->count); |
d6d020e9 | 5409 | |
0cff784a PZ |
5410 | if (!regs) |
5411 | return; | |
5412 | ||
6c7e550f | 5413 | if (!is_sampling_event(event)) |
7b4b6658 | 5414 | return; |
d6d020e9 | 5415 | |
5d81e5cf AV |
5416 | if ((event->attr.sample_type & PERF_SAMPLE_PERIOD) && !event->attr.freq) { |
5417 | data->period = nr; | |
5418 | return perf_swevent_overflow(event, 1, data, regs); | |
5419 | } else | |
5420 | data->period = event->hw.last_period; | |
5421 | ||
0cff784a | 5422 | if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq) |
a8b0ca17 | 5423 | return perf_swevent_overflow(event, 1, data, regs); |
0cff784a | 5424 | |
e7850595 | 5425 | if (local64_add_negative(nr, &hwc->period_left)) |
7b4b6658 | 5426 | return; |
df1a132b | 5427 | |
a8b0ca17 | 5428 | perf_swevent_overflow(event, 0, data, regs); |
d6d020e9 PZ |
5429 | } |
5430 | ||
f5ffe02e FW |
5431 | static int perf_exclude_event(struct perf_event *event, |
5432 | struct pt_regs *regs) | |
5433 | { | |
a4eaf7f1 | 5434 | if (event->hw.state & PERF_HES_STOPPED) |
91b2f482 | 5435 | return 1; |
a4eaf7f1 | 5436 | |
f5ffe02e FW |
5437 | if (regs) { |
5438 | if (event->attr.exclude_user && user_mode(regs)) | |
5439 | return 1; | |
5440 | ||
5441 | if (event->attr.exclude_kernel && !user_mode(regs)) | |
5442 | return 1; | |
5443 | } | |
5444 | ||
5445 | return 0; | |
5446 | } | |
5447 | ||
cdd6c482 | 5448 | static int perf_swevent_match(struct perf_event *event, |
1c432d89 | 5449 | enum perf_type_id type, |
6fb2915d LZ |
5450 | u32 event_id, |
5451 | struct perf_sample_data *data, | |
5452 | struct pt_regs *regs) | |
15dbf27c | 5453 | { |
cdd6c482 | 5454 | if (event->attr.type != type) |
a21ca2ca | 5455 | return 0; |
f5ffe02e | 5456 | |
cdd6c482 | 5457 | if (event->attr.config != event_id) |
15dbf27c PZ |
5458 | return 0; |
5459 | ||
f5ffe02e FW |
5460 | if (perf_exclude_event(event, regs)) |
5461 | return 0; | |
15dbf27c PZ |
5462 | |
5463 | return 1; | |
5464 | } | |
5465 | ||
76e1d904 FW |
5466 | static inline u64 swevent_hash(u64 type, u32 event_id) |
5467 | { | |
5468 | u64 val = event_id | (type << 32); | |
5469 | ||
5470 | return hash_64(val, SWEVENT_HLIST_BITS); | |
5471 | } | |
5472 | ||
49f135ed FW |
5473 | static inline struct hlist_head * |
5474 | __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id) | |
76e1d904 | 5475 | { |
49f135ed FW |
5476 | u64 hash = swevent_hash(type, event_id); |
5477 | ||
5478 | return &hlist->heads[hash]; | |
5479 | } | |
76e1d904 | 5480 | |
49f135ed FW |
5481 | /* For the read side: events when they trigger */ |
5482 | static inline struct hlist_head * | |
b28ab83c | 5483 | find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id) |
49f135ed FW |
5484 | { |
5485 | struct swevent_hlist *hlist; | |
76e1d904 | 5486 | |
b28ab83c | 5487 | hlist = rcu_dereference(swhash->swevent_hlist); |
76e1d904 FW |
5488 | if (!hlist) |
5489 | return NULL; | |
5490 | ||
49f135ed FW |
5491 | return __find_swevent_head(hlist, type, event_id); |
5492 | } | |
5493 | ||
5494 | /* For the event head insertion and removal in the hlist */ | |
5495 | static inline struct hlist_head * | |
b28ab83c | 5496 | find_swevent_head(struct swevent_htable *swhash, struct perf_event *event) |
49f135ed FW |
5497 | { |
5498 | struct swevent_hlist *hlist; | |
5499 | u32 event_id = event->attr.config; | |
5500 | u64 type = event->attr.type; | |
5501 | ||
5502 | /* | |
5503 | * Event scheduling is always serialized against hlist allocation | |
5504 | * and release. Which makes the protected version suitable here. | |
5505 | * The context lock guarantees that. | |
5506 | */ | |
b28ab83c | 5507 | hlist = rcu_dereference_protected(swhash->swevent_hlist, |
49f135ed FW |
5508 | lockdep_is_held(&event->ctx->lock)); |
5509 | if (!hlist) | |
5510 | return NULL; | |
5511 | ||
5512 | return __find_swevent_head(hlist, type, event_id); | |
76e1d904 FW |
5513 | } |
5514 | ||
5515 | static void do_perf_sw_event(enum perf_type_id type, u32 event_id, | |
a8b0ca17 | 5516 | u64 nr, |
76e1d904 FW |
5517 | struct perf_sample_data *data, |
5518 | struct pt_regs *regs) | |
15dbf27c | 5519 | { |
b28ab83c | 5520 | struct swevent_htable *swhash = &__get_cpu_var(swevent_htable); |
cdd6c482 | 5521 | struct perf_event *event; |
76e1d904 | 5522 | struct hlist_head *head; |
15dbf27c | 5523 | |
76e1d904 | 5524 | rcu_read_lock(); |
b28ab83c | 5525 | head = find_swevent_head_rcu(swhash, type, event_id); |
76e1d904 FW |
5526 | if (!head) |
5527 | goto end; | |
5528 | ||
b67bfe0d | 5529 | hlist_for_each_entry_rcu(event, head, hlist_entry) { |
6fb2915d | 5530 | if (perf_swevent_match(event, type, event_id, data, regs)) |
a8b0ca17 | 5531 | perf_swevent_event(event, nr, data, regs); |
15dbf27c | 5532 | } |
76e1d904 FW |
5533 | end: |
5534 | rcu_read_unlock(); | |
15dbf27c PZ |
5535 | } |
5536 | ||
4ed7c92d | 5537 | int perf_swevent_get_recursion_context(void) |
96f6d444 | 5538 | { |
b28ab83c | 5539 | struct swevent_htable *swhash = &__get_cpu_var(swevent_htable); |
96f6d444 | 5540 | |
b28ab83c | 5541 | return get_recursion_context(swhash->recursion); |
96f6d444 | 5542 | } |
645e8cc0 | 5543 | EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context); |
96f6d444 | 5544 | |
fa9f90be | 5545 | inline void perf_swevent_put_recursion_context(int rctx) |
15dbf27c | 5546 | { |
b28ab83c | 5547 | struct swevent_htable *swhash = &__get_cpu_var(swevent_htable); |
927c7a9e | 5548 | |
b28ab83c | 5549 | put_recursion_context(swhash->recursion, rctx); |
ce71b9df | 5550 | } |
15dbf27c | 5551 | |
a8b0ca17 | 5552 | void __perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr) |
b8e83514 | 5553 | { |
a4234bfc | 5554 | struct perf_sample_data data; |
4ed7c92d PZ |
5555 | int rctx; |
5556 | ||
1c024eca | 5557 | preempt_disable_notrace(); |
4ed7c92d PZ |
5558 | rctx = perf_swevent_get_recursion_context(); |
5559 | if (rctx < 0) | |
5560 | return; | |
a4234bfc | 5561 | |
fd0d000b | 5562 | perf_sample_data_init(&data, addr, 0); |
92bf309a | 5563 | |
a8b0ca17 | 5564 | do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, &data, regs); |
4ed7c92d PZ |
5565 | |
5566 | perf_swevent_put_recursion_context(rctx); | |
1c024eca | 5567 | preempt_enable_notrace(); |
b8e83514 PZ |
5568 | } |
5569 | ||
cdd6c482 | 5570 | static void perf_swevent_read(struct perf_event *event) |
15dbf27c | 5571 | { |
15dbf27c PZ |
5572 | } |
5573 | ||
a4eaf7f1 | 5574 | static int perf_swevent_add(struct perf_event *event, int flags) |
15dbf27c | 5575 | { |
b28ab83c | 5576 | struct swevent_htable *swhash = &__get_cpu_var(swevent_htable); |
cdd6c482 | 5577 | struct hw_perf_event *hwc = &event->hw; |
76e1d904 FW |
5578 | struct hlist_head *head; |
5579 | ||
6c7e550f | 5580 | if (is_sampling_event(event)) { |
7b4b6658 | 5581 | hwc->last_period = hwc->sample_period; |
cdd6c482 | 5582 | perf_swevent_set_period(event); |
7b4b6658 | 5583 | } |
76e1d904 | 5584 | |
a4eaf7f1 PZ |
5585 | hwc->state = !(flags & PERF_EF_START); |
5586 | ||
b28ab83c | 5587 | head = find_swevent_head(swhash, event); |
76e1d904 FW |
5588 | if (WARN_ON_ONCE(!head)) |
5589 | return -EINVAL; | |
5590 | ||
5591 | hlist_add_head_rcu(&event->hlist_entry, head); | |
5592 | ||
15dbf27c PZ |
5593 | return 0; |
5594 | } | |
5595 | ||
a4eaf7f1 | 5596 | static void perf_swevent_del(struct perf_event *event, int flags) |
15dbf27c | 5597 | { |
76e1d904 | 5598 | hlist_del_rcu(&event->hlist_entry); |
15dbf27c PZ |
5599 | } |
5600 | ||
a4eaf7f1 | 5601 | static void perf_swevent_start(struct perf_event *event, int flags) |
5c92d124 | 5602 | { |
a4eaf7f1 | 5603 | event->hw.state = 0; |
d6d020e9 | 5604 | } |
aa9c4c0f | 5605 | |
a4eaf7f1 | 5606 | static void perf_swevent_stop(struct perf_event *event, int flags) |
d6d020e9 | 5607 | { |
a4eaf7f1 | 5608 | event->hw.state = PERF_HES_STOPPED; |
bae43c99 IM |
5609 | } |
5610 | ||
49f135ed FW |
5611 | /* Deref the hlist from the update side */ |
5612 | static inline struct swevent_hlist * | |
b28ab83c | 5613 | swevent_hlist_deref(struct swevent_htable *swhash) |
49f135ed | 5614 | { |
b28ab83c PZ |
5615 | return rcu_dereference_protected(swhash->swevent_hlist, |
5616 | lockdep_is_held(&swhash->hlist_mutex)); | |
49f135ed FW |
5617 | } |
5618 | ||
b28ab83c | 5619 | static void swevent_hlist_release(struct swevent_htable *swhash) |
76e1d904 | 5620 | { |
b28ab83c | 5621 | struct swevent_hlist *hlist = swevent_hlist_deref(swhash); |
76e1d904 | 5622 | |
49f135ed | 5623 | if (!hlist) |
76e1d904 FW |
5624 | return; |
5625 | ||
b28ab83c | 5626 | rcu_assign_pointer(swhash->swevent_hlist, NULL); |
fa4bbc4c | 5627 | kfree_rcu(hlist, rcu_head); |
76e1d904 FW |
5628 | } |
5629 | ||
5630 | static void swevent_hlist_put_cpu(struct perf_event *event, int cpu) | |
5631 | { | |
b28ab83c | 5632 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
76e1d904 | 5633 | |
b28ab83c | 5634 | mutex_lock(&swhash->hlist_mutex); |
76e1d904 | 5635 | |
b28ab83c PZ |
5636 | if (!--swhash->hlist_refcount) |
5637 | swevent_hlist_release(swhash); | |
76e1d904 | 5638 | |
b28ab83c | 5639 | mutex_unlock(&swhash->hlist_mutex); |
76e1d904 FW |
5640 | } |
5641 | ||
5642 | static void swevent_hlist_put(struct perf_event *event) | |
5643 | { | |
5644 | int cpu; | |
5645 | ||
5646 | if (event->cpu != -1) { | |
5647 | swevent_hlist_put_cpu(event, event->cpu); | |
5648 | return; | |
5649 | } | |
5650 | ||
5651 | for_each_possible_cpu(cpu) | |
5652 | swevent_hlist_put_cpu(event, cpu); | |
5653 | } | |
5654 | ||
5655 | static int swevent_hlist_get_cpu(struct perf_event *event, int cpu) | |
5656 | { | |
b28ab83c | 5657 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
76e1d904 FW |
5658 | int err = 0; |
5659 | ||
b28ab83c | 5660 | mutex_lock(&swhash->hlist_mutex); |
76e1d904 | 5661 | |
b28ab83c | 5662 | if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) { |
76e1d904 FW |
5663 | struct swevent_hlist *hlist; |
5664 | ||
5665 | hlist = kzalloc(sizeof(*hlist), GFP_KERNEL); | |
5666 | if (!hlist) { | |
5667 | err = -ENOMEM; | |
5668 | goto exit; | |
5669 | } | |
b28ab83c | 5670 | rcu_assign_pointer(swhash->swevent_hlist, hlist); |
76e1d904 | 5671 | } |
b28ab83c | 5672 | swhash->hlist_refcount++; |
9ed6060d | 5673 | exit: |
b28ab83c | 5674 | mutex_unlock(&swhash->hlist_mutex); |
76e1d904 FW |
5675 | |
5676 | return err; | |
5677 | } | |
5678 | ||
5679 | static int swevent_hlist_get(struct perf_event *event) | |
5680 | { | |
5681 | int err; | |
5682 | int cpu, failed_cpu; | |
5683 | ||
5684 | if (event->cpu != -1) | |
5685 | return swevent_hlist_get_cpu(event, event->cpu); | |
5686 | ||
5687 | get_online_cpus(); | |
5688 | for_each_possible_cpu(cpu) { | |
5689 | err = swevent_hlist_get_cpu(event, cpu); | |
5690 | if (err) { | |
5691 | failed_cpu = cpu; | |
5692 | goto fail; | |
5693 | } | |
5694 | } | |
5695 | put_online_cpus(); | |
5696 | ||
5697 | return 0; | |
9ed6060d | 5698 | fail: |
76e1d904 FW |
5699 | for_each_possible_cpu(cpu) { |
5700 | if (cpu == failed_cpu) | |
5701 | break; | |
5702 | swevent_hlist_put_cpu(event, cpu); | |
5703 | } | |
5704 | ||
5705 | put_online_cpus(); | |
5706 | return err; | |
5707 | } | |
5708 | ||
c5905afb | 5709 | struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX]; |
95476b64 | 5710 | |
b0a873eb PZ |
5711 | static void sw_perf_event_destroy(struct perf_event *event) |
5712 | { | |
5713 | u64 event_id = event->attr.config; | |
95476b64 | 5714 | |
b0a873eb PZ |
5715 | WARN_ON(event->parent); |
5716 | ||
c5905afb | 5717 | static_key_slow_dec(&perf_swevent_enabled[event_id]); |
b0a873eb PZ |
5718 | swevent_hlist_put(event); |
5719 | } | |
5720 | ||
5721 | static int perf_swevent_init(struct perf_event *event) | |
5722 | { | |
8176cced | 5723 | u64 event_id = event->attr.config; |
b0a873eb PZ |
5724 | |
5725 | if (event->attr.type != PERF_TYPE_SOFTWARE) | |
5726 | return -ENOENT; | |
5727 | ||
2481c5fa SE |
5728 | /* |
5729 | * no branch sampling for software events | |
5730 | */ | |
5731 | if (has_branch_stack(event)) | |
5732 | return -EOPNOTSUPP; | |
5733 | ||
b0a873eb PZ |
5734 | switch (event_id) { |
5735 | case PERF_COUNT_SW_CPU_CLOCK: | |
5736 | case PERF_COUNT_SW_TASK_CLOCK: | |
5737 | return -ENOENT; | |
5738 | ||
5739 | default: | |
5740 | break; | |
5741 | } | |
5742 | ||
ce677831 | 5743 | if (event_id >= PERF_COUNT_SW_MAX) |
b0a873eb PZ |
5744 | return -ENOENT; |
5745 | ||
5746 | if (!event->parent) { | |
5747 | int err; | |
5748 | ||
5749 | err = swevent_hlist_get(event); | |
5750 | if (err) | |
5751 | return err; | |
5752 | ||
c5905afb | 5753 | static_key_slow_inc(&perf_swevent_enabled[event_id]); |
b0a873eb PZ |
5754 | event->destroy = sw_perf_event_destroy; |
5755 | } | |
5756 | ||
5757 | return 0; | |
5758 | } | |
5759 | ||
35edc2a5 PZ |
5760 | static int perf_swevent_event_idx(struct perf_event *event) |
5761 | { | |
5762 | return 0; | |
5763 | } | |
5764 | ||
b0a873eb | 5765 | static struct pmu perf_swevent = { |
89a1e187 | 5766 | .task_ctx_nr = perf_sw_context, |
95476b64 | 5767 | |
b0a873eb | 5768 | .event_init = perf_swevent_init, |
a4eaf7f1 PZ |
5769 | .add = perf_swevent_add, |
5770 | .del = perf_swevent_del, | |
5771 | .start = perf_swevent_start, | |
5772 | .stop = perf_swevent_stop, | |
1c024eca | 5773 | .read = perf_swevent_read, |
35edc2a5 PZ |
5774 | |
5775 | .event_idx = perf_swevent_event_idx, | |
1c024eca PZ |
5776 | }; |
5777 | ||
b0a873eb PZ |
5778 | #ifdef CONFIG_EVENT_TRACING |
5779 | ||
1c024eca PZ |
5780 | static int perf_tp_filter_match(struct perf_event *event, |
5781 | struct perf_sample_data *data) | |
5782 | { | |
5783 | void *record = data->raw->data; | |
5784 | ||
5785 | if (likely(!event->filter) || filter_match_preds(event->filter, record)) | |
5786 | return 1; | |
5787 | return 0; | |
5788 | } | |
5789 | ||
5790 | static int perf_tp_event_match(struct perf_event *event, | |
5791 | struct perf_sample_data *data, | |
5792 | struct pt_regs *regs) | |
5793 | { | |
a0f7d0f7 FW |
5794 | if (event->hw.state & PERF_HES_STOPPED) |
5795 | return 0; | |
580d607c PZ |
5796 | /* |
5797 | * All tracepoints are from kernel-space. | |
5798 | */ | |
5799 | if (event->attr.exclude_kernel) | |
1c024eca PZ |
5800 | return 0; |
5801 | ||
5802 | if (!perf_tp_filter_match(event, data)) | |
5803 | return 0; | |
5804 | ||
5805 | return 1; | |
5806 | } | |
5807 | ||
5808 | void perf_tp_event(u64 addr, u64 count, void *record, int entry_size, | |
e6dab5ff AV |
5809 | struct pt_regs *regs, struct hlist_head *head, int rctx, |
5810 | struct task_struct *task) | |
95476b64 FW |
5811 | { |
5812 | struct perf_sample_data data; | |
1c024eca | 5813 | struct perf_event *event; |
1c024eca | 5814 | |
95476b64 FW |
5815 | struct perf_raw_record raw = { |
5816 | .size = entry_size, | |
5817 | .data = record, | |
5818 | }; | |
5819 | ||
fd0d000b | 5820 | perf_sample_data_init(&data, addr, 0); |
95476b64 FW |
5821 | data.raw = &raw; |
5822 | ||
b67bfe0d | 5823 | hlist_for_each_entry_rcu(event, head, hlist_entry) { |
1c024eca | 5824 | if (perf_tp_event_match(event, &data, regs)) |
a8b0ca17 | 5825 | perf_swevent_event(event, count, &data, regs); |
4f41c013 | 5826 | } |
ecc55f84 | 5827 | |
e6dab5ff AV |
5828 | /* |
5829 | * If we got specified a target task, also iterate its context and | |
5830 | * deliver this event there too. | |
5831 | */ | |
5832 | if (task && task != current) { | |
5833 | struct perf_event_context *ctx; | |
5834 | struct trace_entry *entry = record; | |
5835 | ||
5836 | rcu_read_lock(); | |
5837 | ctx = rcu_dereference(task->perf_event_ctxp[perf_sw_context]); | |
5838 | if (!ctx) | |
5839 | goto unlock; | |
5840 | ||
5841 | list_for_each_entry_rcu(event, &ctx->event_list, event_entry) { | |
5842 | if (event->attr.type != PERF_TYPE_TRACEPOINT) | |
5843 | continue; | |
5844 | if (event->attr.config != entry->type) | |
5845 | continue; | |
5846 | if (perf_tp_event_match(event, &data, regs)) | |
5847 | perf_swevent_event(event, count, &data, regs); | |
5848 | } | |
5849 | unlock: | |
5850 | rcu_read_unlock(); | |
5851 | } | |
5852 | ||
ecc55f84 | 5853 | perf_swevent_put_recursion_context(rctx); |
95476b64 FW |
5854 | } |
5855 | EXPORT_SYMBOL_GPL(perf_tp_event); | |
5856 | ||
cdd6c482 | 5857 | static void tp_perf_event_destroy(struct perf_event *event) |
e077df4f | 5858 | { |
1c024eca | 5859 | perf_trace_destroy(event); |
e077df4f PZ |
5860 | } |
5861 | ||
b0a873eb | 5862 | static int perf_tp_event_init(struct perf_event *event) |
e077df4f | 5863 | { |
76e1d904 FW |
5864 | int err; |
5865 | ||
b0a873eb PZ |
5866 | if (event->attr.type != PERF_TYPE_TRACEPOINT) |
5867 | return -ENOENT; | |
5868 | ||
2481c5fa SE |
5869 | /* |
5870 | * no branch sampling for tracepoint events | |
5871 | */ | |
5872 | if (has_branch_stack(event)) | |
5873 | return -EOPNOTSUPP; | |
5874 | ||
1c024eca PZ |
5875 | err = perf_trace_init(event); |
5876 | if (err) | |
b0a873eb | 5877 | return err; |
e077df4f | 5878 | |
cdd6c482 | 5879 | event->destroy = tp_perf_event_destroy; |
e077df4f | 5880 | |
b0a873eb PZ |
5881 | return 0; |
5882 | } | |
5883 | ||
5884 | static struct pmu perf_tracepoint = { | |
89a1e187 PZ |
5885 | .task_ctx_nr = perf_sw_context, |
5886 | ||
b0a873eb | 5887 | .event_init = perf_tp_event_init, |
a4eaf7f1 PZ |
5888 | .add = perf_trace_add, |
5889 | .del = perf_trace_del, | |
5890 | .start = perf_swevent_start, | |
5891 | .stop = perf_swevent_stop, | |
b0a873eb | 5892 | .read = perf_swevent_read, |
35edc2a5 PZ |
5893 | |
5894 | .event_idx = perf_swevent_event_idx, | |
b0a873eb PZ |
5895 | }; |
5896 | ||
5897 | static inline void perf_tp_register(void) | |
5898 | { | |
2e80a82a | 5899 | perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT); |
e077df4f | 5900 | } |
6fb2915d LZ |
5901 | |
5902 | static int perf_event_set_filter(struct perf_event *event, void __user *arg) | |
5903 | { | |
5904 | char *filter_str; | |
5905 | int ret; | |
5906 | ||
5907 | if (event->attr.type != PERF_TYPE_TRACEPOINT) | |
5908 | return -EINVAL; | |
5909 | ||
5910 | filter_str = strndup_user(arg, PAGE_SIZE); | |
5911 | if (IS_ERR(filter_str)) | |
5912 | return PTR_ERR(filter_str); | |
5913 | ||
5914 | ret = ftrace_profile_set_filter(event, event->attr.config, filter_str); | |
5915 | ||
5916 | kfree(filter_str); | |
5917 | return ret; | |
5918 | } | |
5919 | ||
5920 | static void perf_event_free_filter(struct perf_event *event) | |
5921 | { | |
5922 | ftrace_profile_free_filter(event); | |
5923 | } | |
5924 | ||
e077df4f | 5925 | #else |
6fb2915d | 5926 | |
b0a873eb | 5927 | static inline void perf_tp_register(void) |
e077df4f | 5928 | { |
e077df4f | 5929 | } |
6fb2915d LZ |
5930 | |
5931 | static int perf_event_set_filter(struct perf_event *event, void __user *arg) | |
5932 | { | |
5933 | return -ENOENT; | |
5934 | } | |
5935 | ||
5936 | static void perf_event_free_filter(struct perf_event *event) | |
5937 | { | |
5938 | } | |
5939 | ||
07b139c8 | 5940 | #endif /* CONFIG_EVENT_TRACING */ |
e077df4f | 5941 | |
24f1e32c | 5942 | #ifdef CONFIG_HAVE_HW_BREAKPOINT |
f5ffe02e | 5943 | void perf_bp_event(struct perf_event *bp, void *data) |
24f1e32c | 5944 | { |
f5ffe02e FW |
5945 | struct perf_sample_data sample; |
5946 | struct pt_regs *regs = data; | |
5947 | ||
fd0d000b | 5948 | perf_sample_data_init(&sample, bp->attr.bp_addr, 0); |
f5ffe02e | 5949 | |
a4eaf7f1 | 5950 | if (!bp->hw.state && !perf_exclude_event(bp, regs)) |
a8b0ca17 | 5951 | perf_swevent_event(bp, 1, &sample, regs); |
24f1e32c FW |
5952 | } |
5953 | #endif | |
5954 | ||
b0a873eb PZ |
5955 | /* |
5956 | * hrtimer based swevent callback | |
5957 | */ | |
f29ac756 | 5958 | |
b0a873eb | 5959 | static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer) |
f29ac756 | 5960 | { |
b0a873eb PZ |
5961 | enum hrtimer_restart ret = HRTIMER_RESTART; |
5962 | struct perf_sample_data data; | |
5963 | struct pt_regs *regs; | |
5964 | struct perf_event *event; | |
5965 | u64 period; | |
f29ac756 | 5966 | |
b0a873eb | 5967 | event = container_of(hrtimer, struct perf_event, hw.hrtimer); |
ba3dd36c PZ |
5968 | |
5969 | if (event->state != PERF_EVENT_STATE_ACTIVE) | |
5970 | return HRTIMER_NORESTART; | |
5971 | ||
b0a873eb | 5972 | event->pmu->read(event); |
f344011c | 5973 | |
fd0d000b | 5974 | perf_sample_data_init(&data, 0, event->hw.last_period); |
b0a873eb PZ |
5975 | regs = get_irq_regs(); |
5976 | ||
5977 | if (regs && !perf_exclude_event(event, regs)) { | |
77aeeebd | 5978 | if (!(event->attr.exclude_idle && is_idle_task(current))) |
33b07b8b | 5979 | if (__perf_event_overflow(event, 1, &data, regs)) |
b0a873eb PZ |
5980 | ret = HRTIMER_NORESTART; |
5981 | } | |
24f1e32c | 5982 | |
b0a873eb PZ |
5983 | period = max_t(u64, 10000, event->hw.sample_period); |
5984 | hrtimer_forward_now(hrtimer, ns_to_ktime(period)); | |
24f1e32c | 5985 | |
b0a873eb | 5986 | return ret; |
f29ac756 PZ |
5987 | } |
5988 | ||
b0a873eb | 5989 | static void perf_swevent_start_hrtimer(struct perf_event *event) |
5c92d124 | 5990 | { |
b0a873eb | 5991 | struct hw_perf_event *hwc = &event->hw; |
5d508e82 FBH |
5992 | s64 period; |
5993 | ||
5994 | if (!is_sampling_event(event)) | |
5995 | return; | |
f5ffe02e | 5996 | |
5d508e82 FBH |
5997 | period = local64_read(&hwc->period_left); |
5998 | if (period) { | |
5999 | if (period < 0) | |
6000 | period = 10000; | |
fa407f35 | 6001 | |
5d508e82 FBH |
6002 | local64_set(&hwc->period_left, 0); |
6003 | } else { | |
6004 | period = max_t(u64, 10000, hwc->sample_period); | |
6005 | } | |
6006 | __hrtimer_start_range_ns(&hwc->hrtimer, | |
b0a873eb | 6007 | ns_to_ktime(period), 0, |
b5ab4cd5 | 6008 | HRTIMER_MODE_REL_PINNED, 0); |
24f1e32c | 6009 | } |
b0a873eb PZ |
6010 | |
6011 | static void perf_swevent_cancel_hrtimer(struct perf_event *event) | |
24f1e32c | 6012 | { |
b0a873eb PZ |
6013 | struct hw_perf_event *hwc = &event->hw; |
6014 | ||
6c7e550f | 6015 | if (is_sampling_event(event)) { |
b0a873eb | 6016 | ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer); |
fa407f35 | 6017 | local64_set(&hwc->period_left, ktime_to_ns(remaining)); |
b0a873eb PZ |
6018 | |
6019 | hrtimer_cancel(&hwc->hrtimer); | |
6020 | } | |
24f1e32c FW |
6021 | } |
6022 | ||
ba3dd36c PZ |
6023 | static void perf_swevent_init_hrtimer(struct perf_event *event) |
6024 | { | |
6025 | struct hw_perf_event *hwc = &event->hw; | |
6026 | ||
6027 | if (!is_sampling_event(event)) | |
6028 | return; | |
6029 | ||
6030 | hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
6031 | hwc->hrtimer.function = perf_swevent_hrtimer; | |
6032 | ||
6033 | /* | |
6034 | * Since hrtimers have a fixed rate, we can do a static freq->period | |
6035 | * mapping and avoid the whole period adjust feedback stuff. | |
6036 | */ | |
6037 | if (event->attr.freq) { | |
6038 | long freq = event->attr.sample_freq; | |
6039 | ||
6040 | event->attr.sample_period = NSEC_PER_SEC / freq; | |
6041 | hwc->sample_period = event->attr.sample_period; | |
6042 | local64_set(&hwc->period_left, hwc->sample_period); | |
778141e3 | 6043 | hwc->last_period = hwc->sample_period; |
ba3dd36c PZ |
6044 | event->attr.freq = 0; |
6045 | } | |
6046 | } | |
6047 | ||
b0a873eb PZ |
6048 | /* |
6049 | * Software event: cpu wall time clock | |
6050 | */ | |
6051 | ||
6052 | static void cpu_clock_event_update(struct perf_event *event) | |
24f1e32c | 6053 | { |
b0a873eb PZ |
6054 | s64 prev; |
6055 | u64 now; | |
6056 | ||
a4eaf7f1 | 6057 | now = local_clock(); |
b0a873eb PZ |
6058 | prev = local64_xchg(&event->hw.prev_count, now); |
6059 | local64_add(now - prev, &event->count); | |
24f1e32c | 6060 | } |
24f1e32c | 6061 | |
a4eaf7f1 | 6062 | static void cpu_clock_event_start(struct perf_event *event, int flags) |
b0a873eb | 6063 | { |
a4eaf7f1 | 6064 | local64_set(&event->hw.prev_count, local_clock()); |
b0a873eb | 6065 | perf_swevent_start_hrtimer(event); |
b0a873eb PZ |
6066 | } |
6067 | ||
a4eaf7f1 | 6068 | static void cpu_clock_event_stop(struct perf_event *event, int flags) |
f29ac756 | 6069 | { |
b0a873eb PZ |
6070 | perf_swevent_cancel_hrtimer(event); |
6071 | cpu_clock_event_update(event); | |
6072 | } | |
f29ac756 | 6073 | |
a4eaf7f1 PZ |
6074 | static int cpu_clock_event_add(struct perf_event *event, int flags) |
6075 | { | |
6076 | if (flags & PERF_EF_START) | |
6077 | cpu_clock_event_start(event, flags); | |
6078 | ||
6079 | return 0; | |
6080 | } | |
6081 | ||
6082 | static void cpu_clock_event_del(struct perf_event *event, int flags) | |
6083 | { | |
6084 | cpu_clock_event_stop(event, flags); | |
6085 | } | |
6086 | ||
b0a873eb PZ |
6087 | static void cpu_clock_event_read(struct perf_event *event) |
6088 | { | |
6089 | cpu_clock_event_update(event); | |
6090 | } | |
f344011c | 6091 | |
b0a873eb PZ |
6092 | static int cpu_clock_event_init(struct perf_event *event) |
6093 | { | |
6094 | if (event->attr.type != PERF_TYPE_SOFTWARE) | |
6095 | return -ENOENT; | |
6096 | ||
6097 | if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK) | |
6098 | return -ENOENT; | |
6099 | ||
2481c5fa SE |
6100 | /* |
6101 | * no branch sampling for software events | |
6102 | */ | |
6103 | if (has_branch_stack(event)) | |
6104 | return -EOPNOTSUPP; | |
6105 | ||
ba3dd36c PZ |
6106 | perf_swevent_init_hrtimer(event); |
6107 | ||
b0a873eb | 6108 | return 0; |
f29ac756 PZ |
6109 | } |
6110 | ||
b0a873eb | 6111 | static struct pmu perf_cpu_clock = { |
89a1e187 PZ |
6112 | .task_ctx_nr = perf_sw_context, |
6113 | ||
b0a873eb | 6114 | .event_init = cpu_clock_event_init, |
a4eaf7f1 PZ |
6115 | .add = cpu_clock_event_add, |
6116 | .del = cpu_clock_event_del, | |
6117 | .start = cpu_clock_event_start, | |
6118 | .stop = cpu_clock_event_stop, | |
b0a873eb | 6119 | .read = cpu_clock_event_read, |
35edc2a5 PZ |
6120 | |
6121 | .event_idx = perf_swevent_event_idx, | |
b0a873eb PZ |
6122 | }; |
6123 | ||
6124 | /* | |
6125 | * Software event: task time clock | |
6126 | */ | |
6127 | ||
6128 | static void task_clock_event_update(struct perf_event *event, u64 now) | |
5c92d124 | 6129 | { |
b0a873eb PZ |
6130 | u64 prev; |
6131 | s64 delta; | |
5c92d124 | 6132 | |
b0a873eb PZ |
6133 | prev = local64_xchg(&event->hw.prev_count, now); |
6134 | delta = now - prev; | |
6135 | local64_add(delta, &event->count); | |
6136 | } | |
5c92d124 | 6137 | |
a4eaf7f1 | 6138 | static void task_clock_event_start(struct perf_event *event, int flags) |
b0a873eb | 6139 | { |
a4eaf7f1 | 6140 | local64_set(&event->hw.prev_count, event->ctx->time); |
b0a873eb | 6141 | perf_swevent_start_hrtimer(event); |
b0a873eb PZ |
6142 | } |
6143 | ||
a4eaf7f1 | 6144 | static void task_clock_event_stop(struct perf_event *event, int flags) |
b0a873eb PZ |
6145 | { |
6146 | perf_swevent_cancel_hrtimer(event); | |
6147 | task_clock_event_update(event, event->ctx->time); | |
a4eaf7f1 PZ |
6148 | } |
6149 | ||
6150 | static int task_clock_event_add(struct perf_event *event, int flags) | |
6151 | { | |
6152 | if (flags & PERF_EF_START) | |
6153 | task_clock_event_start(event, flags); | |
b0a873eb | 6154 | |
a4eaf7f1 PZ |
6155 | return 0; |
6156 | } | |
6157 | ||
6158 | static void task_clock_event_del(struct perf_event *event, int flags) | |
6159 | { | |
6160 | task_clock_event_stop(event, PERF_EF_UPDATE); | |
b0a873eb PZ |
6161 | } |
6162 | ||
6163 | static void task_clock_event_read(struct perf_event *event) | |
6164 | { | |
768a06e2 PZ |
6165 | u64 now = perf_clock(); |
6166 | u64 delta = now - event->ctx->timestamp; | |
6167 | u64 time = event->ctx->time + delta; | |
b0a873eb PZ |
6168 | |
6169 | task_clock_event_update(event, time); | |
6170 | } | |
6171 | ||
6172 | static int task_clock_event_init(struct perf_event *event) | |
6fb2915d | 6173 | { |
b0a873eb PZ |
6174 | if (event->attr.type != PERF_TYPE_SOFTWARE) |
6175 | return -ENOENT; | |
6176 | ||
6177 | if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK) | |
6178 | return -ENOENT; | |
6179 | ||
2481c5fa SE |
6180 | /* |
6181 | * no branch sampling for software events | |
6182 | */ | |
6183 | if (has_branch_stack(event)) | |
6184 | return -EOPNOTSUPP; | |
6185 | ||
ba3dd36c PZ |
6186 | perf_swevent_init_hrtimer(event); |
6187 | ||
b0a873eb | 6188 | return 0; |
6fb2915d LZ |
6189 | } |
6190 | ||
b0a873eb | 6191 | static struct pmu perf_task_clock = { |
89a1e187 PZ |
6192 | .task_ctx_nr = perf_sw_context, |
6193 | ||
b0a873eb | 6194 | .event_init = task_clock_event_init, |
a4eaf7f1 PZ |
6195 | .add = task_clock_event_add, |
6196 | .del = task_clock_event_del, | |
6197 | .start = task_clock_event_start, | |
6198 | .stop = task_clock_event_stop, | |
b0a873eb | 6199 | .read = task_clock_event_read, |
35edc2a5 PZ |
6200 | |
6201 | .event_idx = perf_swevent_event_idx, | |
b0a873eb | 6202 | }; |
6fb2915d | 6203 | |
ad5133b7 | 6204 | static void perf_pmu_nop_void(struct pmu *pmu) |
e077df4f | 6205 | { |
e077df4f | 6206 | } |
6fb2915d | 6207 | |
ad5133b7 | 6208 | static int perf_pmu_nop_int(struct pmu *pmu) |
6fb2915d | 6209 | { |
ad5133b7 | 6210 | return 0; |
6fb2915d LZ |
6211 | } |
6212 | ||
ad5133b7 | 6213 | static void perf_pmu_start_txn(struct pmu *pmu) |
6fb2915d | 6214 | { |
ad5133b7 | 6215 | perf_pmu_disable(pmu); |
6fb2915d LZ |
6216 | } |
6217 | ||
ad5133b7 PZ |
6218 | static int perf_pmu_commit_txn(struct pmu *pmu) |
6219 | { | |
6220 | perf_pmu_enable(pmu); | |
6221 | return 0; | |
6222 | } | |
e077df4f | 6223 | |
ad5133b7 | 6224 | static void perf_pmu_cancel_txn(struct pmu *pmu) |
24f1e32c | 6225 | { |
ad5133b7 | 6226 | perf_pmu_enable(pmu); |
24f1e32c FW |
6227 | } |
6228 | ||
35edc2a5 PZ |
6229 | static int perf_event_idx_default(struct perf_event *event) |
6230 | { | |
6231 | return event->hw.idx + 1; | |
6232 | } | |
6233 | ||
8dc85d54 PZ |
6234 | /* |
6235 | * Ensures all contexts with the same task_ctx_nr have the same | |
6236 | * pmu_cpu_context too. | |
6237 | */ | |
6238 | static void *find_pmu_context(int ctxn) | |
24f1e32c | 6239 | { |
8dc85d54 | 6240 | struct pmu *pmu; |
b326e956 | 6241 | |
8dc85d54 PZ |
6242 | if (ctxn < 0) |
6243 | return NULL; | |
24f1e32c | 6244 | |
8dc85d54 PZ |
6245 | list_for_each_entry(pmu, &pmus, entry) { |
6246 | if (pmu->task_ctx_nr == ctxn) | |
6247 | return pmu->pmu_cpu_context; | |
6248 | } | |
24f1e32c | 6249 | |
8dc85d54 | 6250 | return NULL; |
24f1e32c FW |
6251 | } |
6252 | ||
51676957 | 6253 | static void update_pmu_context(struct pmu *pmu, struct pmu *old_pmu) |
24f1e32c | 6254 | { |
51676957 PZ |
6255 | int cpu; |
6256 | ||
6257 | for_each_possible_cpu(cpu) { | |
6258 | struct perf_cpu_context *cpuctx; | |
6259 | ||
6260 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); | |
6261 | ||
3f1f3320 PZ |
6262 | if (cpuctx->unique_pmu == old_pmu) |
6263 | cpuctx->unique_pmu = pmu; | |
51676957 PZ |
6264 | } |
6265 | } | |
6266 | ||
6267 | static void free_pmu_context(struct pmu *pmu) | |
6268 | { | |
6269 | struct pmu *i; | |
f5ffe02e | 6270 | |
8dc85d54 | 6271 | mutex_lock(&pmus_lock); |
0475f9ea | 6272 | /* |
8dc85d54 | 6273 | * Like a real lame refcount. |
0475f9ea | 6274 | */ |
51676957 PZ |
6275 | list_for_each_entry(i, &pmus, entry) { |
6276 | if (i->pmu_cpu_context == pmu->pmu_cpu_context) { | |
6277 | update_pmu_context(i, pmu); | |
8dc85d54 | 6278 | goto out; |
51676957 | 6279 | } |
8dc85d54 | 6280 | } |
d6d020e9 | 6281 | |
51676957 | 6282 | free_percpu(pmu->pmu_cpu_context); |
8dc85d54 PZ |
6283 | out: |
6284 | mutex_unlock(&pmus_lock); | |
24f1e32c | 6285 | } |
2e80a82a | 6286 | static struct idr pmu_idr; |
d6d020e9 | 6287 | |
abe43400 PZ |
6288 | static ssize_t |
6289 | type_show(struct device *dev, struct device_attribute *attr, char *page) | |
6290 | { | |
6291 | struct pmu *pmu = dev_get_drvdata(dev); | |
6292 | ||
6293 | return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->type); | |
6294 | } | |
6295 | ||
62b85639 SE |
6296 | static ssize_t |
6297 | perf_event_mux_interval_ms_show(struct device *dev, | |
6298 | struct device_attribute *attr, | |
6299 | char *page) | |
6300 | { | |
6301 | struct pmu *pmu = dev_get_drvdata(dev); | |
6302 | ||
6303 | return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->hrtimer_interval_ms); | |
6304 | } | |
6305 | ||
6306 | static ssize_t | |
6307 | perf_event_mux_interval_ms_store(struct device *dev, | |
6308 | struct device_attribute *attr, | |
6309 | const char *buf, size_t count) | |
6310 | { | |
6311 | struct pmu *pmu = dev_get_drvdata(dev); | |
6312 | int timer, cpu, ret; | |
6313 | ||
6314 | ret = kstrtoint(buf, 0, &timer); | |
6315 | if (ret) | |
6316 | return ret; | |
6317 | ||
6318 | if (timer < 1) | |
6319 | return -EINVAL; | |
6320 | ||
6321 | /* same value, noting to do */ | |
6322 | if (timer == pmu->hrtimer_interval_ms) | |
6323 | return count; | |
6324 | ||
6325 | pmu->hrtimer_interval_ms = timer; | |
6326 | ||
6327 | /* update all cpuctx for this PMU */ | |
6328 | for_each_possible_cpu(cpu) { | |
6329 | struct perf_cpu_context *cpuctx; | |
6330 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); | |
6331 | cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * timer); | |
6332 | ||
6333 | if (hrtimer_active(&cpuctx->hrtimer)) | |
6334 | hrtimer_forward_now(&cpuctx->hrtimer, cpuctx->hrtimer_interval); | |
6335 | } | |
6336 | ||
6337 | return count; | |
6338 | } | |
6339 | ||
abe43400 | 6340 | static struct device_attribute pmu_dev_attrs[] = { |
62b85639 SE |
6341 | __ATTR_RO(type), |
6342 | __ATTR_RW(perf_event_mux_interval_ms), | |
6343 | __ATTR_NULL, | |
abe43400 PZ |
6344 | }; |
6345 | ||
6346 | static int pmu_bus_running; | |
6347 | static struct bus_type pmu_bus = { | |
6348 | .name = "event_source", | |
6349 | .dev_attrs = pmu_dev_attrs, | |
6350 | }; | |
6351 | ||
6352 | static void pmu_dev_release(struct device *dev) | |
6353 | { | |
6354 | kfree(dev); | |
6355 | } | |
6356 | ||
6357 | static int pmu_dev_alloc(struct pmu *pmu) | |
6358 | { | |
6359 | int ret = -ENOMEM; | |
6360 | ||
6361 | pmu->dev = kzalloc(sizeof(struct device), GFP_KERNEL); | |
6362 | if (!pmu->dev) | |
6363 | goto out; | |
6364 | ||
0c9d42ed | 6365 | pmu->dev->groups = pmu->attr_groups; |
abe43400 PZ |
6366 | device_initialize(pmu->dev); |
6367 | ret = dev_set_name(pmu->dev, "%s", pmu->name); | |
6368 | if (ret) | |
6369 | goto free_dev; | |
6370 | ||
6371 | dev_set_drvdata(pmu->dev, pmu); | |
6372 | pmu->dev->bus = &pmu_bus; | |
6373 | pmu->dev->release = pmu_dev_release; | |
6374 | ret = device_add(pmu->dev); | |
6375 | if (ret) | |
6376 | goto free_dev; | |
6377 | ||
6378 | out: | |
6379 | return ret; | |
6380 | ||
6381 | free_dev: | |
6382 | put_device(pmu->dev); | |
6383 | goto out; | |
6384 | } | |
6385 | ||
547e9fd7 | 6386 | static struct lock_class_key cpuctx_mutex; |
facc4307 | 6387 | static struct lock_class_key cpuctx_lock; |
547e9fd7 | 6388 | |
03d8e80b | 6389 | int perf_pmu_register(struct pmu *pmu, const char *name, int type) |
24f1e32c | 6390 | { |
108b02cf | 6391 | int cpu, ret; |
24f1e32c | 6392 | |
b0a873eb | 6393 | mutex_lock(&pmus_lock); |
33696fc0 PZ |
6394 | ret = -ENOMEM; |
6395 | pmu->pmu_disable_count = alloc_percpu(int); | |
6396 | if (!pmu->pmu_disable_count) | |
6397 | goto unlock; | |
f29ac756 | 6398 | |
2e80a82a PZ |
6399 | pmu->type = -1; |
6400 | if (!name) | |
6401 | goto skip_type; | |
6402 | pmu->name = name; | |
6403 | ||
6404 | if (type < 0) { | |
0e9c3be2 TH |
6405 | type = idr_alloc(&pmu_idr, pmu, PERF_TYPE_MAX, 0, GFP_KERNEL); |
6406 | if (type < 0) { | |
6407 | ret = type; | |
2e80a82a PZ |
6408 | goto free_pdc; |
6409 | } | |
6410 | } | |
6411 | pmu->type = type; | |
6412 | ||
abe43400 PZ |
6413 | if (pmu_bus_running) { |
6414 | ret = pmu_dev_alloc(pmu); | |
6415 | if (ret) | |
6416 | goto free_idr; | |
6417 | } | |
6418 | ||
2e80a82a | 6419 | skip_type: |
8dc85d54 PZ |
6420 | pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr); |
6421 | if (pmu->pmu_cpu_context) | |
6422 | goto got_cpu_context; | |
f29ac756 | 6423 | |
c4814202 | 6424 | ret = -ENOMEM; |
108b02cf PZ |
6425 | pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context); |
6426 | if (!pmu->pmu_cpu_context) | |
abe43400 | 6427 | goto free_dev; |
f344011c | 6428 | |
108b02cf PZ |
6429 | for_each_possible_cpu(cpu) { |
6430 | struct perf_cpu_context *cpuctx; | |
6431 | ||
6432 | cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu); | |
eb184479 | 6433 | __perf_event_init_context(&cpuctx->ctx); |
547e9fd7 | 6434 | lockdep_set_class(&cpuctx->ctx.mutex, &cpuctx_mutex); |
facc4307 | 6435 | lockdep_set_class(&cpuctx->ctx.lock, &cpuctx_lock); |
b04243ef | 6436 | cpuctx->ctx.type = cpu_context; |
108b02cf | 6437 | cpuctx->ctx.pmu = pmu; |
9e630205 SE |
6438 | |
6439 | __perf_cpu_hrtimer_init(cpuctx, cpu); | |
6440 | ||
e9d2b064 | 6441 | INIT_LIST_HEAD(&cpuctx->rotation_list); |
3f1f3320 | 6442 | cpuctx->unique_pmu = pmu; |
108b02cf | 6443 | } |
76e1d904 | 6444 | |
8dc85d54 | 6445 | got_cpu_context: |
ad5133b7 PZ |
6446 | if (!pmu->start_txn) { |
6447 | if (pmu->pmu_enable) { | |
6448 | /* | |
6449 | * If we have pmu_enable/pmu_disable calls, install | |
6450 | * transaction stubs that use that to try and batch | |
6451 | * hardware accesses. | |
6452 | */ | |
6453 | pmu->start_txn = perf_pmu_start_txn; | |
6454 | pmu->commit_txn = perf_pmu_commit_txn; | |
6455 | pmu->cancel_txn = perf_pmu_cancel_txn; | |
6456 | } else { | |
6457 | pmu->start_txn = perf_pmu_nop_void; | |
6458 | pmu->commit_txn = perf_pmu_nop_int; | |
6459 | pmu->cancel_txn = perf_pmu_nop_void; | |
f344011c | 6460 | } |
5c92d124 | 6461 | } |
15dbf27c | 6462 | |
ad5133b7 PZ |
6463 | if (!pmu->pmu_enable) { |
6464 | pmu->pmu_enable = perf_pmu_nop_void; | |
6465 | pmu->pmu_disable = perf_pmu_nop_void; | |
6466 | } | |
6467 | ||
35edc2a5 PZ |
6468 | if (!pmu->event_idx) |
6469 | pmu->event_idx = perf_event_idx_default; | |
6470 | ||
b0a873eb | 6471 | list_add_rcu(&pmu->entry, &pmus); |
33696fc0 PZ |
6472 | ret = 0; |
6473 | unlock: | |
b0a873eb PZ |
6474 | mutex_unlock(&pmus_lock); |
6475 | ||
33696fc0 | 6476 | return ret; |
108b02cf | 6477 | |
abe43400 PZ |
6478 | free_dev: |
6479 | device_del(pmu->dev); | |
6480 | put_device(pmu->dev); | |
6481 | ||
2e80a82a PZ |
6482 | free_idr: |
6483 | if (pmu->type >= PERF_TYPE_MAX) | |
6484 | idr_remove(&pmu_idr, pmu->type); | |
6485 | ||
108b02cf PZ |
6486 | free_pdc: |
6487 | free_percpu(pmu->pmu_disable_count); | |
6488 | goto unlock; | |
f29ac756 PZ |
6489 | } |
6490 | ||
b0a873eb | 6491 | void perf_pmu_unregister(struct pmu *pmu) |
5c92d124 | 6492 | { |
b0a873eb PZ |
6493 | mutex_lock(&pmus_lock); |
6494 | list_del_rcu(&pmu->entry); | |
6495 | mutex_unlock(&pmus_lock); | |
5c92d124 | 6496 | |
0475f9ea | 6497 | /* |
cde8e884 PZ |
6498 | * We dereference the pmu list under both SRCU and regular RCU, so |
6499 | * synchronize against both of those. | |
0475f9ea | 6500 | */ |
b0a873eb | 6501 | synchronize_srcu(&pmus_srcu); |
cde8e884 | 6502 | synchronize_rcu(); |
d6d020e9 | 6503 | |
33696fc0 | 6504 | free_percpu(pmu->pmu_disable_count); |
2e80a82a PZ |
6505 | if (pmu->type >= PERF_TYPE_MAX) |
6506 | idr_remove(&pmu_idr, pmu->type); | |
abe43400 PZ |
6507 | device_del(pmu->dev); |
6508 | put_device(pmu->dev); | |
51676957 | 6509 | free_pmu_context(pmu); |
b0a873eb | 6510 | } |
d6d020e9 | 6511 | |
b0a873eb PZ |
6512 | struct pmu *perf_init_event(struct perf_event *event) |
6513 | { | |
6514 | struct pmu *pmu = NULL; | |
6515 | int idx; | |
940c5b29 | 6516 | int ret; |
b0a873eb PZ |
6517 | |
6518 | idx = srcu_read_lock(&pmus_srcu); | |
2e80a82a PZ |
6519 | |
6520 | rcu_read_lock(); | |
6521 | pmu = idr_find(&pmu_idr, event->attr.type); | |
6522 | rcu_read_unlock(); | |
940c5b29 | 6523 | if (pmu) { |
7e5b2a01 | 6524 | event->pmu = pmu; |
940c5b29 LM |
6525 | ret = pmu->event_init(event); |
6526 | if (ret) | |
6527 | pmu = ERR_PTR(ret); | |
2e80a82a | 6528 | goto unlock; |
940c5b29 | 6529 | } |
2e80a82a | 6530 | |
b0a873eb | 6531 | list_for_each_entry_rcu(pmu, &pmus, entry) { |
7e5b2a01 | 6532 | event->pmu = pmu; |
940c5b29 | 6533 | ret = pmu->event_init(event); |
b0a873eb | 6534 | if (!ret) |
e5f4d339 | 6535 | goto unlock; |
76e1d904 | 6536 | |
b0a873eb PZ |
6537 | if (ret != -ENOENT) { |
6538 | pmu = ERR_PTR(ret); | |
e5f4d339 | 6539 | goto unlock; |
f344011c | 6540 | } |
5c92d124 | 6541 | } |
e5f4d339 PZ |
6542 | pmu = ERR_PTR(-ENOENT); |
6543 | unlock: | |
b0a873eb | 6544 | srcu_read_unlock(&pmus_srcu, idx); |
15dbf27c | 6545 | |
4aeb0b42 | 6546 | return pmu; |
5c92d124 IM |
6547 | } |
6548 | ||
4beb31f3 FW |
6549 | static void account_event_cpu(struct perf_event *event, int cpu) |
6550 | { | |
6551 | if (event->parent) | |
6552 | return; | |
6553 | ||
6554 | if (has_branch_stack(event)) { | |
6555 | if (!(event->attach_state & PERF_ATTACH_TASK)) | |
6556 | atomic_inc(&per_cpu(perf_branch_stack_events, cpu)); | |
6557 | } | |
6558 | if (is_cgroup_event(event)) | |
6559 | atomic_inc(&per_cpu(perf_cgroup_events, cpu)); | |
6560 | } | |
6561 | ||
766d6c07 FW |
6562 | static void account_event(struct perf_event *event) |
6563 | { | |
4beb31f3 FW |
6564 | if (event->parent) |
6565 | return; | |
6566 | ||
766d6c07 FW |
6567 | if (event->attach_state & PERF_ATTACH_TASK) |
6568 | static_key_slow_inc(&perf_sched_events.key); | |
6569 | if (event->attr.mmap || event->attr.mmap_data) | |
6570 | atomic_inc(&nr_mmap_events); | |
6571 | if (event->attr.comm) | |
6572 | atomic_inc(&nr_comm_events); | |
6573 | if (event->attr.task) | |
6574 | atomic_inc(&nr_task_events); | |
948b26b6 FW |
6575 | if (event->attr.freq) { |
6576 | if (atomic_inc_return(&nr_freq_events) == 1) | |
6577 | tick_nohz_full_kick_all(); | |
6578 | } | |
4beb31f3 | 6579 | if (has_branch_stack(event)) |
766d6c07 | 6580 | static_key_slow_inc(&perf_sched_events.key); |
4beb31f3 | 6581 | if (is_cgroup_event(event)) |
766d6c07 | 6582 | static_key_slow_inc(&perf_sched_events.key); |
4beb31f3 FW |
6583 | |
6584 | account_event_cpu(event, event->cpu); | |
766d6c07 FW |
6585 | } |
6586 | ||
0793a61d | 6587 | /* |
cdd6c482 | 6588 | * Allocate and initialize a event structure |
0793a61d | 6589 | */ |
cdd6c482 | 6590 | static struct perf_event * |
c3f00c70 | 6591 | perf_event_alloc(struct perf_event_attr *attr, int cpu, |
d580ff86 PZ |
6592 | struct task_struct *task, |
6593 | struct perf_event *group_leader, | |
6594 | struct perf_event *parent_event, | |
4dc0da86 AK |
6595 | perf_overflow_handler_t overflow_handler, |
6596 | void *context) | |
0793a61d | 6597 | { |
51b0fe39 | 6598 | struct pmu *pmu; |
cdd6c482 IM |
6599 | struct perf_event *event; |
6600 | struct hw_perf_event *hwc; | |
90983b16 | 6601 | long err = -EINVAL; |
0793a61d | 6602 | |
66832eb4 ON |
6603 | if ((unsigned)cpu >= nr_cpu_ids) { |
6604 | if (!task || cpu != -1) | |
6605 | return ERR_PTR(-EINVAL); | |
6606 | } | |
6607 | ||
c3f00c70 | 6608 | event = kzalloc(sizeof(*event), GFP_KERNEL); |
cdd6c482 | 6609 | if (!event) |
d5d2bc0d | 6610 | return ERR_PTR(-ENOMEM); |
0793a61d | 6611 | |
04289bb9 | 6612 | /* |
cdd6c482 | 6613 | * Single events are their own group leaders, with an |
04289bb9 IM |
6614 | * empty sibling list: |
6615 | */ | |
6616 | if (!group_leader) | |
cdd6c482 | 6617 | group_leader = event; |
04289bb9 | 6618 | |
cdd6c482 IM |
6619 | mutex_init(&event->child_mutex); |
6620 | INIT_LIST_HEAD(&event->child_list); | |
fccc714b | 6621 | |
cdd6c482 IM |
6622 | INIT_LIST_HEAD(&event->group_entry); |
6623 | INIT_LIST_HEAD(&event->event_entry); | |
6624 | INIT_LIST_HEAD(&event->sibling_list); | |
10c6db11 PZ |
6625 | INIT_LIST_HEAD(&event->rb_entry); |
6626 | ||
cdd6c482 | 6627 | init_waitqueue_head(&event->waitq); |
e360adbe | 6628 | init_irq_work(&event->pending, perf_pending_event); |
0793a61d | 6629 | |
cdd6c482 | 6630 | mutex_init(&event->mmap_mutex); |
7b732a75 | 6631 | |
a6fa941d | 6632 | atomic_long_set(&event->refcount, 1); |
cdd6c482 IM |
6633 | event->cpu = cpu; |
6634 | event->attr = *attr; | |
6635 | event->group_leader = group_leader; | |
6636 | event->pmu = NULL; | |
cdd6c482 | 6637 | event->oncpu = -1; |
a96bbc16 | 6638 | |
cdd6c482 | 6639 | event->parent = parent_event; |
b84fbc9f | 6640 | |
17cf22c3 | 6641 | event->ns = get_pid_ns(task_active_pid_ns(current)); |
cdd6c482 | 6642 | event->id = atomic64_inc_return(&perf_event_id); |
a96bbc16 | 6643 | |
cdd6c482 | 6644 | event->state = PERF_EVENT_STATE_INACTIVE; |
329d876d | 6645 | |
d580ff86 PZ |
6646 | if (task) { |
6647 | event->attach_state = PERF_ATTACH_TASK; | |
f22c1bb6 ON |
6648 | |
6649 | if (attr->type == PERF_TYPE_TRACEPOINT) | |
6650 | event->hw.tp_target = task; | |
d580ff86 PZ |
6651 | #ifdef CONFIG_HAVE_HW_BREAKPOINT |
6652 | /* | |
6653 | * hw_breakpoint is a bit difficult here.. | |
6654 | */ | |
f22c1bb6 | 6655 | else if (attr->type == PERF_TYPE_BREAKPOINT) |
d580ff86 PZ |
6656 | event->hw.bp_target = task; |
6657 | #endif | |
6658 | } | |
6659 | ||
4dc0da86 | 6660 | if (!overflow_handler && parent_event) { |
b326e956 | 6661 | overflow_handler = parent_event->overflow_handler; |
4dc0da86 AK |
6662 | context = parent_event->overflow_handler_context; |
6663 | } | |
66832eb4 | 6664 | |
b326e956 | 6665 | event->overflow_handler = overflow_handler; |
4dc0da86 | 6666 | event->overflow_handler_context = context; |
97eaf530 | 6667 | |
0231bb53 | 6668 | perf_event__state_init(event); |
a86ed508 | 6669 | |
4aeb0b42 | 6670 | pmu = NULL; |
b8e83514 | 6671 | |
cdd6c482 | 6672 | hwc = &event->hw; |
bd2b5b12 | 6673 | hwc->sample_period = attr->sample_period; |
0d48696f | 6674 | if (attr->freq && attr->sample_freq) |
bd2b5b12 | 6675 | hwc->sample_period = 1; |
eced1dfc | 6676 | hwc->last_period = hwc->sample_period; |
bd2b5b12 | 6677 | |
e7850595 | 6678 | local64_set(&hwc->period_left, hwc->sample_period); |
60db5e09 | 6679 | |
2023b359 | 6680 | /* |
cdd6c482 | 6681 | * we currently do not support PERF_FORMAT_GROUP on inherited events |
2023b359 | 6682 | */ |
3dab77fb | 6683 | if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP)) |
90983b16 | 6684 | goto err_ns; |
2023b359 | 6685 | |
b0a873eb | 6686 | pmu = perf_init_event(event); |
4aeb0b42 | 6687 | if (!pmu) |
90983b16 FW |
6688 | goto err_ns; |
6689 | else if (IS_ERR(pmu)) { | |
4aeb0b42 | 6690 | err = PTR_ERR(pmu); |
90983b16 | 6691 | goto err_ns; |
621a01ea | 6692 | } |
d5d2bc0d | 6693 | |
cdd6c482 | 6694 | if (!event->parent) { |
927c7a9e FW |
6695 | if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) { |
6696 | err = get_callchain_buffers(); | |
90983b16 FW |
6697 | if (err) |
6698 | goto err_pmu; | |
d010b332 | 6699 | } |
f344011c | 6700 | } |
9ee318a7 | 6701 | |
cdd6c482 | 6702 | return event; |
90983b16 FW |
6703 | |
6704 | err_pmu: | |
6705 | if (event->destroy) | |
6706 | event->destroy(event); | |
6707 | err_ns: | |
6708 | if (event->ns) | |
6709 | put_pid_ns(event->ns); | |
6710 | kfree(event); | |
6711 | ||
6712 | return ERR_PTR(err); | |
0793a61d TG |
6713 | } |
6714 | ||
cdd6c482 IM |
6715 | static int perf_copy_attr(struct perf_event_attr __user *uattr, |
6716 | struct perf_event_attr *attr) | |
974802ea | 6717 | { |
974802ea | 6718 | u32 size; |
cdf8073d | 6719 | int ret; |
974802ea PZ |
6720 | |
6721 | if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0)) | |
6722 | return -EFAULT; | |
6723 | ||
6724 | /* | |
6725 | * zero the full structure, so that a short copy will be nice. | |
6726 | */ | |
6727 | memset(attr, 0, sizeof(*attr)); | |
6728 | ||
6729 | ret = get_user(size, &uattr->size); | |
6730 | if (ret) | |
6731 | return ret; | |
6732 | ||
6733 | if (size > PAGE_SIZE) /* silly large */ | |
6734 | goto err_size; | |
6735 | ||
6736 | if (!size) /* abi compat */ | |
6737 | size = PERF_ATTR_SIZE_VER0; | |
6738 | ||
6739 | if (size < PERF_ATTR_SIZE_VER0) | |
6740 | goto err_size; | |
6741 | ||
6742 | /* | |
6743 | * If we're handed a bigger struct than we know of, | |
cdf8073d IS |
6744 | * ensure all the unknown bits are 0 - i.e. new |
6745 | * user-space does not rely on any kernel feature | |
6746 | * extensions we dont know about yet. | |
974802ea PZ |
6747 | */ |
6748 | if (size > sizeof(*attr)) { | |
cdf8073d IS |
6749 | unsigned char __user *addr; |
6750 | unsigned char __user *end; | |
6751 | unsigned char val; | |
974802ea | 6752 | |
cdf8073d IS |
6753 | addr = (void __user *)uattr + sizeof(*attr); |
6754 | end = (void __user *)uattr + size; | |
974802ea | 6755 | |
cdf8073d | 6756 | for (; addr < end; addr++) { |
974802ea PZ |
6757 | ret = get_user(val, addr); |
6758 | if (ret) | |
6759 | return ret; | |
6760 | if (val) | |
6761 | goto err_size; | |
6762 | } | |
b3e62e35 | 6763 | size = sizeof(*attr); |
974802ea PZ |
6764 | } |
6765 | ||
6766 | ret = copy_from_user(attr, uattr, size); | |
6767 | if (ret) | |
6768 | return -EFAULT; | |
6769 | ||
cd757645 | 6770 | if (attr->__reserved_1) |
974802ea PZ |
6771 | return -EINVAL; |
6772 | ||
6773 | if (attr->sample_type & ~(PERF_SAMPLE_MAX-1)) | |
6774 | return -EINVAL; | |
6775 | ||
6776 | if (attr->read_format & ~(PERF_FORMAT_MAX-1)) | |
6777 | return -EINVAL; | |
6778 | ||
bce38cd5 SE |
6779 | if (attr->sample_type & PERF_SAMPLE_BRANCH_STACK) { |
6780 | u64 mask = attr->branch_sample_type; | |
6781 | ||
6782 | /* only using defined bits */ | |
6783 | if (mask & ~(PERF_SAMPLE_BRANCH_MAX-1)) | |
6784 | return -EINVAL; | |
6785 | ||
6786 | /* at least one branch bit must be set */ | |
6787 | if (!(mask & ~PERF_SAMPLE_BRANCH_PLM_ALL)) | |
6788 | return -EINVAL; | |
6789 | ||
bce38cd5 SE |
6790 | /* propagate priv level, when not set for branch */ |
6791 | if (!(mask & PERF_SAMPLE_BRANCH_PLM_ALL)) { | |
6792 | ||
6793 | /* exclude_kernel checked on syscall entry */ | |
6794 | if (!attr->exclude_kernel) | |
6795 | mask |= PERF_SAMPLE_BRANCH_KERNEL; | |
6796 | ||
6797 | if (!attr->exclude_user) | |
6798 | mask |= PERF_SAMPLE_BRANCH_USER; | |
6799 | ||
6800 | if (!attr->exclude_hv) | |
6801 | mask |= PERF_SAMPLE_BRANCH_HV; | |
6802 | /* | |
6803 | * adjust user setting (for HW filter setup) | |
6804 | */ | |
6805 | attr->branch_sample_type = mask; | |
6806 | } | |
e712209a SE |
6807 | /* privileged levels capture (kernel, hv): check permissions */ |
6808 | if ((mask & PERF_SAMPLE_BRANCH_PERM_PLM) | |
2b923c8f SE |
6809 | && perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN)) |
6810 | return -EACCES; | |
bce38cd5 | 6811 | } |
4018994f | 6812 | |
c5ebcedb | 6813 | if (attr->sample_type & PERF_SAMPLE_REGS_USER) { |
4018994f | 6814 | ret = perf_reg_validate(attr->sample_regs_user); |
c5ebcedb JO |
6815 | if (ret) |
6816 | return ret; | |
6817 | } | |
6818 | ||
6819 | if (attr->sample_type & PERF_SAMPLE_STACK_USER) { | |
6820 | if (!arch_perf_have_user_stack_dump()) | |
6821 | return -ENOSYS; | |
6822 | ||
6823 | /* | |
6824 | * We have __u32 type for the size, but so far | |
6825 | * we can only use __u16 as maximum due to the | |
6826 | * __u16 sample size limit. | |
6827 | */ | |
6828 | if (attr->sample_stack_user >= USHRT_MAX) | |
6829 | ret = -EINVAL; | |
6830 | else if (!IS_ALIGNED(attr->sample_stack_user, sizeof(u64))) | |
6831 | ret = -EINVAL; | |
6832 | } | |
4018994f | 6833 | |
974802ea PZ |
6834 | out: |
6835 | return ret; | |
6836 | ||
6837 | err_size: | |
6838 | put_user(sizeof(*attr), &uattr->size); | |
6839 | ret = -E2BIG; | |
6840 | goto out; | |
6841 | } | |
6842 | ||
ac9721f3 PZ |
6843 | static int |
6844 | perf_event_set_output(struct perf_event *event, struct perf_event *output_event) | |
a4be7c27 | 6845 | { |
76369139 | 6846 | struct ring_buffer *rb = NULL, *old_rb = NULL; |
a4be7c27 PZ |
6847 | int ret = -EINVAL; |
6848 | ||
ac9721f3 | 6849 | if (!output_event) |
a4be7c27 PZ |
6850 | goto set; |
6851 | ||
ac9721f3 PZ |
6852 | /* don't allow circular references */ |
6853 | if (event == output_event) | |
a4be7c27 PZ |
6854 | goto out; |
6855 | ||
0f139300 PZ |
6856 | /* |
6857 | * Don't allow cross-cpu buffers | |
6858 | */ | |
6859 | if (output_event->cpu != event->cpu) | |
6860 | goto out; | |
6861 | ||
6862 | /* | |
76369139 | 6863 | * If its not a per-cpu rb, it must be the same task. |
0f139300 PZ |
6864 | */ |
6865 | if (output_event->cpu == -1 && output_event->ctx != event->ctx) | |
6866 | goto out; | |
6867 | ||
a4be7c27 | 6868 | set: |
cdd6c482 | 6869 | mutex_lock(&event->mmap_mutex); |
ac9721f3 PZ |
6870 | /* Can't redirect output if we've got an active mmap() */ |
6871 | if (atomic_read(&event->mmap_count)) | |
6872 | goto unlock; | |
a4be7c27 | 6873 | |
9bb5d40c PZ |
6874 | old_rb = event->rb; |
6875 | ||
ac9721f3 | 6876 | if (output_event) { |
76369139 FW |
6877 | /* get the rb we want to redirect to */ |
6878 | rb = ring_buffer_get(output_event); | |
6879 | if (!rb) | |
ac9721f3 | 6880 | goto unlock; |
a4be7c27 PZ |
6881 | } |
6882 | ||
10c6db11 PZ |
6883 | if (old_rb) |
6884 | ring_buffer_detach(event, old_rb); | |
9bb5d40c PZ |
6885 | |
6886 | if (rb) | |
6887 | ring_buffer_attach(event, rb); | |
6888 | ||
6889 | rcu_assign_pointer(event->rb, rb); | |
6890 | ||
6891 | if (old_rb) { | |
6892 | ring_buffer_put(old_rb); | |
6893 | /* | |
6894 | * Since we detached before setting the new rb, so that we | |
6895 | * could attach the new rb, we could have missed a wakeup. | |
6896 | * Provide it now. | |
6897 | */ | |
6898 | wake_up_all(&event->waitq); | |
6899 | } | |
6900 | ||
a4be7c27 | 6901 | ret = 0; |
ac9721f3 PZ |
6902 | unlock: |
6903 | mutex_unlock(&event->mmap_mutex); | |
6904 | ||
a4be7c27 | 6905 | out: |
a4be7c27 PZ |
6906 | return ret; |
6907 | } | |
6908 | ||
0793a61d | 6909 | /** |
cdd6c482 | 6910 | * sys_perf_event_open - open a performance event, associate it to a task/cpu |
9f66a381 | 6911 | * |
cdd6c482 | 6912 | * @attr_uptr: event_id type attributes for monitoring/sampling |
0793a61d | 6913 | * @pid: target pid |
9f66a381 | 6914 | * @cpu: target cpu |
cdd6c482 | 6915 | * @group_fd: group leader event fd |
0793a61d | 6916 | */ |
cdd6c482 IM |
6917 | SYSCALL_DEFINE5(perf_event_open, |
6918 | struct perf_event_attr __user *, attr_uptr, | |
2743a5b0 | 6919 | pid_t, pid, int, cpu, int, group_fd, unsigned long, flags) |
0793a61d | 6920 | { |
b04243ef PZ |
6921 | struct perf_event *group_leader = NULL, *output_event = NULL; |
6922 | struct perf_event *event, *sibling; | |
cdd6c482 IM |
6923 | struct perf_event_attr attr; |
6924 | struct perf_event_context *ctx; | |
6925 | struct file *event_file = NULL; | |
2903ff01 | 6926 | struct fd group = {NULL, 0}; |
38a81da2 | 6927 | struct task_struct *task = NULL; |
89a1e187 | 6928 | struct pmu *pmu; |
ea635c64 | 6929 | int event_fd; |
b04243ef | 6930 | int move_group = 0; |
dc86cabe | 6931 | int err; |
0793a61d | 6932 | |
2743a5b0 | 6933 | /* for future expandability... */ |
e5d1367f | 6934 | if (flags & ~PERF_FLAG_ALL) |
2743a5b0 PM |
6935 | return -EINVAL; |
6936 | ||
dc86cabe IM |
6937 | err = perf_copy_attr(attr_uptr, &attr); |
6938 | if (err) | |
6939 | return err; | |
eab656ae | 6940 | |
0764771d PZ |
6941 | if (!attr.exclude_kernel) { |
6942 | if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN)) | |
6943 | return -EACCES; | |
6944 | } | |
6945 | ||
df58ab24 | 6946 | if (attr.freq) { |
cdd6c482 | 6947 | if (attr.sample_freq > sysctl_perf_event_sample_rate) |
df58ab24 PZ |
6948 | return -EINVAL; |
6949 | } | |
6950 | ||
e5d1367f SE |
6951 | /* |
6952 | * In cgroup mode, the pid argument is used to pass the fd | |
6953 | * opened to the cgroup directory in cgroupfs. The cpu argument | |
6954 | * designates the cpu on which to monitor threads from that | |
6955 | * cgroup. | |
6956 | */ | |
6957 | if ((flags & PERF_FLAG_PID_CGROUP) && (pid == -1 || cpu == -1)) | |
6958 | return -EINVAL; | |
6959 | ||
ab72a702 | 6960 | event_fd = get_unused_fd(); |
ea635c64 AV |
6961 | if (event_fd < 0) |
6962 | return event_fd; | |
6963 | ||
ac9721f3 | 6964 | if (group_fd != -1) { |
2903ff01 AV |
6965 | err = perf_fget_light(group_fd, &group); |
6966 | if (err) | |
d14b12d7 | 6967 | goto err_fd; |
2903ff01 | 6968 | group_leader = group.file->private_data; |
ac9721f3 PZ |
6969 | if (flags & PERF_FLAG_FD_OUTPUT) |
6970 | output_event = group_leader; | |
6971 | if (flags & PERF_FLAG_FD_NO_GROUP) | |
6972 | group_leader = NULL; | |
6973 | } | |
6974 | ||
e5d1367f | 6975 | if (pid != -1 && !(flags & PERF_FLAG_PID_CGROUP)) { |
c6be5a5c PZ |
6976 | task = find_lively_task_by_vpid(pid); |
6977 | if (IS_ERR(task)) { | |
6978 | err = PTR_ERR(task); | |
6979 | goto err_group_fd; | |
6980 | } | |
6981 | } | |
6982 | ||
fbfc623f YZ |
6983 | get_online_cpus(); |
6984 | ||
4dc0da86 AK |
6985 | event = perf_event_alloc(&attr, cpu, task, group_leader, NULL, |
6986 | NULL, NULL); | |
d14b12d7 SE |
6987 | if (IS_ERR(event)) { |
6988 | err = PTR_ERR(event); | |
c6be5a5c | 6989 | goto err_task; |
d14b12d7 SE |
6990 | } |
6991 | ||
e5d1367f SE |
6992 | if (flags & PERF_FLAG_PID_CGROUP) { |
6993 | err = perf_cgroup_connect(pid, event, &attr, group_leader); | |
766d6c07 FW |
6994 | if (err) { |
6995 | __free_event(event); | |
6996 | goto err_task; | |
6997 | } | |
e5d1367f SE |
6998 | } |
6999 | ||
766d6c07 FW |
7000 | account_event(event); |
7001 | ||
89a1e187 PZ |
7002 | /* |
7003 | * Special case software events and allow them to be part of | |
7004 | * any hardware group. | |
7005 | */ | |
7006 | pmu = event->pmu; | |
b04243ef PZ |
7007 | |
7008 | if (group_leader && | |
7009 | (is_software_event(event) != is_software_event(group_leader))) { | |
7010 | if (is_software_event(event)) { | |
7011 | /* | |
7012 | * If event and group_leader are not both a software | |
7013 | * event, and event is, then group leader is not. | |
7014 | * | |
7015 | * Allow the addition of software events to !software | |
7016 | * groups, this is safe because software events never | |
7017 | * fail to schedule. | |
7018 | */ | |
7019 | pmu = group_leader->pmu; | |
7020 | } else if (is_software_event(group_leader) && | |
7021 | (group_leader->group_flags & PERF_GROUP_SOFTWARE)) { | |
7022 | /* | |
7023 | * In case the group is a pure software group, and we | |
7024 | * try to add a hardware event, move the whole group to | |
7025 | * the hardware context. | |
7026 | */ | |
7027 | move_group = 1; | |
7028 | } | |
7029 | } | |
89a1e187 PZ |
7030 | |
7031 | /* | |
7032 | * Get the target context (task or percpu): | |
7033 | */ | |
e2d37cd2 | 7034 | ctx = find_get_context(pmu, task, event->cpu); |
89a1e187 PZ |
7035 | if (IS_ERR(ctx)) { |
7036 | err = PTR_ERR(ctx); | |
c6be5a5c | 7037 | goto err_alloc; |
89a1e187 PZ |
7038 | } |
7039 | ||
fd1edb3a PZ |
7040 | if (task) { |
7041 | put_task_struct(task); | |
7042 | task = NULL; | |
7043 | } | |
7044 | ||
ccff286d | 7045 | /* |
cdd6c482 | 7046 | * Look up the group leader (we will attach this event to it): |
04289bb9 | 7047 | */ |
ac9721f3 | 7048 | if (group_leader) { |
dc86cabe | 7049 | err = -EINVAL; |
04289bb9 | 7050 | |
04289bb9 | 7051 | /* |
ccff286d IM |
7052 | * Do not allow a recursive hierarchy (this new sibling |
7053 | * becoming part of another group-sibling): | |
7054 | */ | |
7055 | if (group_leader->group_leader != group_leader) | |
c3f00c70 | 7056 | goto err_context; |
ccff286d IM |
7057 | /* |
7058 | * Do not allow to attach to a group in a different | |
7059 | * task or CPU context: | |
04289bb9 | 7060 | */ |
b04243ef PZ |
7061 | if (move_group) { |
7062 | if (group_leader->ctx->type != ctx->type) | |
7063 | goto err_context; | |
7064 | } else { | |
7065 | if (group_leader->ctx != ctx) | |
7066 | goto err_context; | |
7067 | } | |
7068 | ||
3b6f9e5c PM |
7069 | /* |
7070 | * Only a group leader can be exclusive or pinned | |
7071 | */ | |
0d48696f | 7072 | if (attr.exclusive || attr.pinned) |
c3f00c70 | 7073 | goto err_context; |
ac9721f3 PZ |
7074 | } |
7075 | ||
7076 | if (output_event) { | |
7077 | err = perf_event_set_output(event, output_event); | |
7078 | if (err) | |
c3f00c70 | 7079 | goto err_context; |
ac9721f3 | 7080 | } |
0793a61d | 7081 | |
ea635c64 AV |
7082 | event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, O_RDWR); |
7083 | if (IS_ERR(event_file)) { | |
7084 | err = PTR_ERR(event_file); | |
c3f00c70 | 7085 | goto err_context; |
ea635c64 | 7086 | } |
9b51f66d | 7087 | |
b04243ef PZ |
7088 | if (move_group) { |
7089 | struct perf_event_context *gctx = group_leader->ctx; | |
7090 | ||
7091 | mutex_lock(&gctx->mutex); | |
fe4b04fa | 7092 | perf_remove_from_context(group_leader); |
0231bb53 JO |
7093 | |
7094 | /* | |
7095 | * Removing from the context ends up with disabled | |
7096 | * event. What we want here is event in the initial | |
7097 | * startup state, ready to be add into new context. | |
7098 | */ | |
7099 | perf_event__state_init(group_leader); | |
b04243ef PZ |
7100 | list_for_each_entry(sibling, &group_leader->sibling_list, |
7101 | group_entry) { | |
fe4b04fa | 7102 | perf_remove_from_context(sibling); |
0231bb53 | 7103 | perf_event__state_init(sibling); |
b04243ef PZ |
7104 | put_ctx(gctx); |
7105 | } | |
7106 | mutex_unlock(&gctx->mutex); | |
7107 | put_ctx(gctx); | |
ea635c64 | 7108 | } |
9b51f66d | 7109 | |
ad3a37de | 7110 | WARN_ON_ONCE(ctx->parent_ctx); |
d859e29f | 7111 | mutex_lock(&ctx->mutex); |
b04243ef PZ |
7112 | |
7113 | if (move_group) { | |
0cda4c02 | 7114 | synchronize_rcu(); |
e2d37cd2 | 7115 | perf_install_in_context(ctx, group_leader, event->cpu); |
b04243ef PZ |
7116 | get_ctx(ctx); |
7117 | list_for_each_entry(sibling, &group_leader->sibling_list, | |
7118 | group_entry) { | |
e2d37cd2 | 7119 | perf_install_in_context(ctx, sibling, event->cpu); |
b04243ef PZ |
7120 | get_ctx(ctx); |
7121 | } | |
7122 | } | |
7123 | ||
e2d37cd2 | 7124 | perf_install_in_context(ctx, event, event->cpu); |
ad3a37de | 7125 | ++ctx->generation; |
fe4b04fa | 7126 | perf_unpin_context(ctx); |
d859e29f | 7127 | mutex_unlock(&ctx->mutex); |
9b51f66d | 7128 | |
fbfc623f YZ |
7129 | put_online_cpus(); |
7130 | ||
cdd6c482 | 7131 | event->owner = current; |
8882135b | 7132 | |
cdd6c482 IM |
7133 | mutex_lock(¤t->perf_event_mutex); |
7134 | list_add_tail(&event->owner_entry, ¤t->perf_event_list); | |
7135 | mutex_unlock(¤t->perf_event_mutex); | |
082ff5a2 | 7136 | |
c320c7b7 ACM |
7137 | /* |
7138 | * Precalculate sample_data sizes | |
7139 | */ | |
7140 | perf_event__header_size(event); | |
6844c09d | 7141 | perf_event__id_header_size(event); |
c320c7b7 | 7142 | |
8a49542c PZ |
7143 | /* |
7144 | * Drop the reference on the group_event after placing the | |
7145 | * new event on the sibling_list. This ensures destruction | |
7146 | * of the group leader will find the pointer to itself in | |
7147 | * perf_group_detach(). | |
7148 | */ | |
2903ff01 | 7149 | fdput(group); |
ea635c64 AV |
7150 | fd_install(event_fd, event_file); |
7151 | return event_fd; | |
0793a61d | 7152 | |
c3f00c70 | 7153 | err_context: |
fe4b04fa | 7154 | perf_unpin_context(ctx); |
ea635c64 | 7155 | put_ctx(ctx); |
c6be5a5c | 7156 | err_alloc: |
ea635c64 | 7157 | free_event(event); |
e7d0bc04 | 7158 | err_task: |
fbfc623f | 7159 | put_online_cpus(); |
e7d0bc04 PZ |
7160 | if (task) |
7161 | put_task_struct(task); | |
89a1e187 | 7162 | err_group_fd: |
2903ff01 | 7163 | fdput(group); |
ea635c64 AV |
7164 | err_fd: |
7165 | put_unused_fd(event_fd); | |
dc86cabe | 7166 | return err; |
0793a61d TG |
7167 | } |
7168 | ||
fb0459d7 AV |
7169 | /** |
7170 | * perf_event_create_kernel_counter | |
7171 | * | |
7172 | * @attr: attributes of the counter to create | |
7173 | * @cpu: cpu in which the counter is bound | |
38a81da2 | 7174 | * @task: task to profile (NULL for percpu) |
fb0459d7 AV |
7175 | */ |
7176 | struct perf_event * | |
7177 | perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu, | |
38a81da2 | 7178 | struct task_struct *task, |
4dc0da86 AK |
7179 | perf_overflow_handler_t overflow_handler, |
7180 | void *context) | |
fb0459d7 | 7181 | { |
fb0459d7 | 7182 | struct perf_event_context *ctx; |
c3f00c70 | 7183 | struct perf_event *event; |
fb0459d7 | 7184 | int err; |
d859e29f | 7185 | |
fb0459d7 AV |
7186 | /* |
7187 | * Get the target context (task or percpu): | |
7188 | */ | |
d859e29f | 7189 | |
4dc0da86 AK |
7190 | event = perf_event_alloc(attr, cpu, task, NULL, NULL, |
7191 | overflow_handler, context); | |
c3f00c70 PZ |
7192 | if (IS_ERR(event)) { |
7193 | err = PTR_ERR(event); | |
7194 | goto err; | |
7195 | } | |
d859e29f | 7196 | |
766d6c07 FW |
7197 | account_event(event); |
7198 | ||
38a81da2 | 7199 | ctx = find_get_context(event->pmu, task, cpu); |
c6567f64 FW |
7200 | if (IS_ERR(ctx)) { |
7201 | err = PTR_ERR(ctx); | |
c3f00c70 | 7202 | goto err_free; |
d859e29f | 7203 | } |
fb0459d7 | 7204 | |
fb0459d7 AV |
7205 | WARN_ON_ONCE(ctx->parent_ctx); |
7206 | mutex_lock(&ctx->mutex); | |
7207 | perf_install_in_context(ctx, event, cpu); | |
7208 | ++ctx->generation; | |
fe4b04fa | 7209 | perf_unpin_context(ctx); |
fb0459d7 AV |
7210 | mutex_unlock(&ctx->mutex); |
7211 | ||
fb0459d7 AV |
7212 | return event; |
7213 | ||
c3f00c70 PZ |
7214 | err_free: |
7215 | free_event(event); | |
7216 | err: | |
c6567f64 | 7217 | return ERR_PTR(err); |
9b51f66d | 7218 | } |
fb0459d7 | 7219 | EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter); |
9b51f66d | 7220 | |
0cda4c02 YZ |
7221 | void perf_pmu_migrate_context(struct pmu *pmu, int src_cpu, int dst_cpu) |
7222 | { | |
7223 | struct perf_event_context *src_ctx; | |
7224 | struct perf_event_context *dst_ctx; | |
7225 | struct perf_event *event, *tmp; | |
7226 | LIST_HEAD(events); | |
7227 | ||
7228 | src_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, src_cpu)->ctx; | |
7229 | dst_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, dst_cpu)->ctx; | |
7230 | ||
7231 | mutex_lock(&src_ctx->mutex); | |
7232 | list_for_each_entry_safe(event, tmp, &src_ctx->event_list, | |
7233 | event_entry) { | |
7234 | perf_remove_from_context(event); | |
9a545de0 | 7235 | unaccount_event_cpu(event, src_cpu); |
0cda4c02 | 7236 | put_ctx(src_ctx); |
9886167d | 7237 | list_add(&event->migrate_entry, &events); |
0cda4c02 YZ |
7238 | } |
7239 | mutex_unlock(&src_ctx->mutex); | |
7240 | ||
7241 | synchronize_rcu(); | |
7242 | ||
7243 | mutex_lock(&dst_ctx->mutex); | |
9886167d PZ |
7244 | list_for_each_entry_safe(event, tmp, &events, migrate_entry) { |
7245 | list_del(&event->migrate_entry); | |
0cda4c02 YZ |
7246 | if (event->state >= PERF_EVENT_STATE_OFF) |
7247 | event->state = PERF_EVENT_STATE_INACTIVE; | |
9a545de0 | 7248 | account_event_cpu(event, dst_cpu); |
0cda4c02 YZ |
7249 | perf_install_in_context(dst_ctx, event, dst_cpu); |
7250 | get_ctx(dst_ctx); | |
7251 | } | |
7252 | mutex_unlock(&dst_ctx->mutex); | |
7253 | } | |
7254 | EXPORT_SYMBOL_GPL(perf_pmu_migrate_context); | |
7255 | ||
cdd6c482 | 7256 | static void sync_child_event(struct perf_event *child_event, |
38b200d6 | 7257 | struct task_struct *child) |
d859e29f | 7258 | { |
cdd6c482 | 7259 | struct perf_event *parent_event = child_event->parent; |
8bc20959 | 7260 | u64 child_val; |
d859e29f | 7261 | |
cdd6c482 IM |
7262 | if (child_event->attr.inherit_stat) |
7263 | perf_event_read_event(child_event, child); | |
38b200d6 | 7264 | |
b5e58793 | 7265 | child_val = perf_event_count(child_event); |
d859e29f PM |
7266 | |
7267 | /* | |
7268 | * Add back the child's count to the parent's count: | |
7269 | */ | |
a6e6dea6 | 7270 | atomic64_add(child_val, &parent_event->child_count); |
cdd6c482 IM |
7271 | atomic64_add(child_event->total_time_enabled, |
7272 | &parent_event->child_total_time_enabled); | |
7273 | atomic64_add(child_event->total_time_running, | |
7274 | &parent_event->child_total_time_running); | |
d859e29f PM |
7275 | |
7276 | /* | |
cdd6c482 | 7277 | * Remove this event from the parent's list |
d859e29f | 7278 | */ |
cdd6c482 IM |
7279 | WARN_ON_ONCE(parent_event->ctx->parent_ctx); |
7280 | mutex_lock(&parent_event->child_mutex); | |
7281 | list_del_init(&child_event->child_list); | |
7282 | mutex_unlock(&parent_event->child_mutex); | |
d859e29f PM |
7283 | |
7284 | /* | |
cdd6c482 | 7285 | * Release the parent event, if this was the last |
d859e29f PM |
7286 | * reference to it. |
7287 | */ | |
a6fa941d | 7288 | put_event(parent_event); |
d859e29f PM |
7289 | } |
7290 | ||
9b51f66d | 7291 | static void |
cdd6c482 IM |
7292 | __perf_event_exit_task(struct perf_event *child_event, |
7293 | struct perf_event_context *child_ctx, | |
38b200d6 | 7294 | struct task_struct *child) |
9b51f66d | 7295 | { |
38b435b1 PZ |
7296 | if (child_event->parent) { |
7297 | raw_spin_lock_irq(&child_ctx->lock); | |
7298 | perf_group_detach(child_event); | |
7299 | raw_spin_unlock_irq(&child_ctx->lock); | |
7300 | } | |
9b51f66d | 7301 | |
fe4b04fa | 7302 | perf_remove_from_context(child_event); |
0cc0c027 | 7303 | |
9b51f66d | 7304 | /* |
38b435b1 | 7305 | * It can happen that the parent exits first, and has events |
9b51f66d | 7306 | * that are still around due to the child reference. These |
38b435b1 | 7307 | * events need to be zapped. |
9b51f66d | 7308 | */ |
38b435b1 | 7309 | if (child_event->parent) { |
cdd6c482 IM |
7310 | sync_child_event(child_event, child); |
7311 | free_event(child_event); | |
4bcf349a | 7312 | } |
9b51f66d IM |
7313 | } |
7314 | ||
8dc85d54 | 7315 | static void perf_event_exit_task_context(struct task_struct *child, int ctxn) |
9b51f66d | 7316 | { |
cdd6c482 IM |
7317 | struct perf_event *child_event, *tmp; |
7318 | struct perf_event_context *child_ctx; | |
a63eaf34 | 7319 | unsigned long flags; |
9b51f66d | 7320 | |
8dc85d54 | 7321 | if (likely(!child->perf_event_ctxp[ctxn])) { |
cdd6c482 | 7322 | perf_event_task(child, NULL, 0); |
9b51f66d | 7323 | return; |
9f498cc5 | 7324 | } |
9b51f66d | 7325 | |
a63eaf34 | 7326 | local_irq_save(flags); |
ad3a37de PM |
7327 | /* |
7328 | * We can't reschedule here because interrupts are disabled, | |
7329 | * and either child is current or it is a task that can't be | |
7330 | * scheduled, so we are now safe from rescheduling changing | |
7331 | * our context. | |
7332 | */ | |
806839b2 | 7333 | child_ctx = rcu_dereference_raw(child->perf_event_ctxp[ctxn]); |
c93f7669 PM |
7334 | |
7335 | /* | |
7336 | * Take the context lock here so that if find_get_context is | |
cdd6c482 | 7337 | * reading child->perf_event_ctxp, we wait until it has |
c93f7669 PM |
7338 | * incremented the context's refcount before we do put_ctx below. |
7339 | */ | |
e625cce1 | 7340 | raw_spin_lock(&child_ctx->lock); |
04dc2dbb | 7341 | task_ctx_sched_out(child_ctx); |
8dc85d54 | 7342 | child->perf_event_ctxp[ctxn] = NULL; |
71a851b4 PZ |
7343 | /* |
7344 | * If this context is a clone; unclone it so it can't get | |
7345 | * swapped to another process while we're removing all | |
cdd6c482 | 7346 | * the events from it. |
71a851b4 PZ |
7347 | */ |
7348 | unclone_ctx(child_ctx); | |
5e942bb3 | 7349 | update_context_time(child_ctx); |
e625cce1 | 7350 | raw_spin_unlock_irqrestore(&child_ctx->lock, flags); |
9f498cc5 PZ |
7351 | |
7352 | /* | |
cdd6c482 IM |
7353 | * Report the task dead after unscheduling the events so that we |
7354 | * won't get any samples after PERF_RECORD_EXIT. We can however still | |
7355 | * get a few PERF_RECORD_READ events. | |
9f498cc5 | 7356 | */ |
cdd6c482 | 7357 | perf_event_task(child, child_ctx, 0); |
a63eaf34 | 7358 | |
66fff224 PZ |
7359 | /* |
7360 | * We can recurse on the same lock type through: | |
7361 | * | |
cdd6c482 IM |
7362 | * __perf_event_exit_task() |
7363 | * sync_child_event() | |
a6fa941d AV |
7364 | * put_event() |
7365 | * mutex_lock(&ctx->mutex) | |
66fff224 PZ |
7366 | * |
7367 | * But since its the parent context it won't be the same instance. | |
7368 | */ | |
a0507c84 | 7369 | mutex_lock(&child_ctx->mutex); |
a63eaf34 | 7370 | |
8bc20959 | 7371 | again: |
889ff015 FW |
7372 | list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups, |
7373 | group_entry) | |
7374 | __perf_event_exit_task(child_event, child_ctx, child); | |
7375 | ||
7376 | list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups, | |
65abc865 | 7377 | group_entry) |
cdd6c482 | 7378 | __perf_event_exit_task(child_event, child_ctx, child); |
8bc20959 PZ |
7379 | |
7380 | /* | |
cdd6c482 | 7381 | * If the last event was a group event, it will have appended all |
8bc20959 PZ |
7382 | * its siblings to the list, but we obtained 'tmp' before that which |
7383 | * will still point to the list head terminating the iteration. | |
7384 | */ | |
889ff015 FW |
7385 | if (!list_empty(&child_ctx->pinned_groups) || |
7386 | !list_empty(&child_ctx->flexible_groups)) | |
8bc20959 | 7387 | goto again; |
a63eaf34 PM |
7388 | |
7389 | mutex_unlock(&child_ctx->mutex); | |
7390 | ||
7391 | put_ctx(child_ctx); | |
9b51f66d IM |
7392 | } |
7393 | ||
8dc85d54 PZ |
7394 | /* |
7395 | * When a child task exits, feed back event values to parent events. | |
7396 | */ | |
7397 | void perf_event_exit_task(struct task_struct *child) | |
7398 | { | |
8882135b | 7399 | struct perf_event *event, *tmp; |
8dc85d54 PZ |
7400 | int ctxn; |
7401 | ||
8882135b PZ |
7402 | mutex_lock(&child->perf_event_mutex); |
7403 | list_for_each_entry_safe(event, tmp, &child->perf_event_list, | |
7404 | owner_entry) { | |
7405 | list_del_init(&event->owner_entry); | |
7406 | ||
7407 | /* | |
7408 | * Ensure the list deletion is visible before we clear | |
7409 | * the owner, closes a race against perf_release() where | |
7410 | * we need to serialize on the owner->perf_event_mutex. | |
7411 | */ | |
7412 | smp_wmb(); | |
7413 | event->owner = NULL; | |
7414 | } | |
7415 | mutex_unlock(&child->perf_event_mutex); | |
7416 | ||
8dc85d54 PZ |
7417 | for_each_task_context_nr(ctxn) |
7418 | perf_event_exit_task_context(child, ctxn); | |
7419 | } | |
7420 | ||
889ff015 FW |
7421 | static void perf_free_event(struct perf_event *event, |
7422 | struct perf_event_context *ctx) | |
7423 | { | |
7424 | struct perf_event *parent = event->parent; | |
7425 | ||
7426 | if (WARN_ON_ONCE(!parent)) | |
7427 | return; | |
7428 | ||
7429 | mutex_lock(&parent->child_mutex); | |
7430 | list_del_init(&event->child_list); | |
7431 | mutex_unlock(&parent->child_mutex); | |
7432 | ||
a6fa941d | 7433 | put_event(parent); |
889ff015 | 7434 | |
8a49542c | 7435 | perf_group_detach(event); |
889ff015 FW |
7436 | list_del_event(event, ctx); |
7437 | free_event(event); | |
7438 | } | |
7439 | ||
bbbee908 PZ |
7440 | /* |
7441 | * free an unexposed, unused context as created by inheritance by | |
8dc85d54 | 7442 | * perf_event_init_task below, used by fork() in case of fail. |
bbbee908 | 7443 | */ |
cdd6c482 | 7444 | void perf_event_free_task(struct task_struct *task) |
bbbee908 | 7445 | { |
8dc85d54 | 7446 | struct perf_event_context *ctx; |
cdd6c482 | 7447 | struct perf_event *event, *tmp; |
8dc85d54 | 7448 | int ctxn; |
bbbee908 | 7449 | |
8dc85d54 PZ |
7450 | for_each_task_context_nr(ctxn) { |
7451 | ctx = task->perf_event_ctxp[ctxn]; | |
7452 | if (!ctx) | |
7453 | continue; | |
bbbee908 | 7454 | |
8dc85d54 | 7455 | mutex_lock(&ctx->mutex); |
bbbee908 | 7456 | again: |
8dc85d54 PZ |
7457 | list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, |
7458 | group_entry) | |
7459 | perf_free_event(event, ctx); | |
bbbee908 | 7460 | |
8dc85d54 PZ |
7461 | list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, |
7462 | group_entry) | |
7463 | perf_free_event(event, ctx); | |
bbbee908 | 7464 | |
8dc85d54 PZ |
7465 | if (!list_empty(&ctx->pinned_groups) || |
7466 | !list_empty(&ctx->flexible_groups)) | |
7467 | goto again; | |
bbbee908 | 7468 | |
8dc85d54 | 7469 | mutex_unlock(&ctx->mutex); |
bbbee908 | 7470 | |
8dc85d54 PZ |
7471 | put_ctx(ctx); |
7472 | } | |
889ff015 FW |
7473 | } |
7474 | ||
4e231c79 PZ |
7475 | void perf_event_delayed_put(struct task_struct *task) |
7476 | { | |
7477 | int ctxn; | |
7478 | ||
7479 | for_each_task_context_nr(ctxn) | |
7480 | WARN_ON_ONCE(task->perf_event_ctxp[ctxn]); | |
7481 | } | |
7482 | ||
97dee4f3 PZ |
7483 | /* |
7484 | * inherit a event from parent task to child task: | |
7485 | */ | |
7486 | static struct perf_event * | |
7487 | inherit_event(struct perf_event *parent_event, | |
7488 | struct task_struct *parent, | |
7489 | struct perf_event_context *parent_ctx, | |
7490 | struct task_struct *child, | |
7491 | struct perf_event *group_leader, | |
7492 | struct perf_event_context *child_ctx) | |
7493 | { | |
7494 | struct perf_event *child_event; | |
cee010ec | 7495 | unsigned long flags; |
97dee4f3 PZ |
7496 | |
7497 | /* | |
7498 | * Instead of creating recursive hierarchies of events, | |
7499 | * we link inherited events back to the original parent, | |
7500 | * which has a filp for sure, which we use as the reference | |
7501 | * count: | |
7502 | */ | |
7503 | if (parent_event->parent) | |
7504 | parent_event = parent_event->parent; | |
7505 | ||
7506 | child_event = perf_event_alloc(&parent_event->attr, | |
7507 | parent_event->cpu, | |
d580ff86 | 7508 | child, |
97dee4f3 | 7509 | group_leader, parent_event, |
4dc0da86 | 7510 | NULL, NULL); |
97dee4f3 PZ |
7511 | if (IS_ERR(child_event)) |
7512 | return child_event; | |
a6fa941d AV |
7513 | |
7514 | if (!atomic_long_inc_not_zero(&parent_event->refcount)) { | |
7515 | free_event(child_event); | |
7516 | return NULL; | |
7517 | } | |
7518 | ||
97dee4f3 PZ |
7519 | get_ctx(child_ctx); |
7520 | ||
7521 | /* | |
7522 | * Make the child state follow the state of the parent event, | |
7523 | * not its attr.disabled bit. We hold the parent's mutex, | |
7524 | * so we won't race with perf_event_{en, dis}able_family. | |
7525 | */ | |
7526 | if (parent_event->state >= PERF_EVENT_STATE_INACTIVE) | |
7527 | child_event->state = PERF_EVENT_STATE_INACTIVE; | |
7528 | else | |
7529 | child_event->state = PERF_EVENT_STATE_OFF; | |
7530 | ||
7531 | if (parent_event->attr.freq) { | |
7532 | u64 sample_period = parent_event->hw.sample_period; | |
7533 | struct hw_perf_event *hwc = &child_event->hw; | |
7534 | ||
7535 | hwc->sample_period = sample_period; | |
7536 | hwc->last_period = sample_period; | |
7537 | ||
7538 | local64_set(&hwc->period_left, sample_period); | |
7539 | } | |
7540 | ||
7541 | child_event->ctx = child_ctx; | |
7542 | child_event->overflow_handler = parent_event->overflow_handler; | |
4dc0da86 AK |
7543 | child_event->overflow_handler_context |
7544 | = parent_event->overflow_handler_context; | |
97dee4f3 | 7545 | |
614b6780 TG |
7546 | /* |
7547 | * Precalculate sample_data sizes | |
7548 | */ | |
7549 | perf_event__header_size(child_event); | |
6844c09d | 7550 | perf_event__id_header_size(child_event); |
614b6780 | 7551 | |
97dee4f3 PZ |
7552 | /* |
7553 | * Link it up in the child's context: | |
7554 | */ | |
cee010ec | 7555 | raw_spin_lock_irqsave(&child_ctx->lock, flags); |
97dee4f3 | 7556 | add_event_to_ctx(child_event, child_ctx); |
cee010ec | 7557 | raw_spin_unlock_irqrestore(&child_ctx->lock, flags); |
97dee4f3 | 7558 | |
97dee4f3 PZ |
7559 | /* |
7560 | * Link this into the parent event's child list | |
7561 | */ | |
7562 | WARN_ON_ONCE(parent_event->ctx->parent_ctx); | |
7563 | mutex_lock(&parent_event->child_mutex); | |
7564 | list_add_tail(&child_event->child_list, &parent_event->child_list); | |
7565 | mutex_unlock(&parent_event->child_mutex); | |
7566 | ||
7567 | return child_event; | |
7568 | } | |
7569 | ||
7570 | static int inherit_group(struct perf_event *parent_event, | |
7571 | struct task_struct *parent, | |
7572 | struct perf_event_context *parent_ctx, | |
7573 | struct task_struct *child, | |
7574 | struct perf_event_context *child_ctx) | |
7575 | { | |
7576 | struct perf_event *leader; | |
7577 | struct perf_event *sub; | |
7578 | struct perf_event *child_ctr; | |
7579 | ||
7580 | leader = inherit_event(parent_event, parent, parent_ctx, | |
7581 | child, NULL, child_ctx); | |
7582 | if (IS_ERR(leader)) | |
7583 | return PTR_ERR(leader); | |
7584 | list_for_each_entry(sub, &parent_event->sibling_list, group_entry) { | |
7585 | child_ctr = inherit_event(sub, parent, parent_ctx, | |
7586 | child, leader, child_ctx); | |
7587 | if (IS_ERR(child_ctr)) | |
7588 | return PTR_ERR(child_ctr); | |
7589 | } | |
7590 | return 0; | |
889ff015 FW |
7591 | } |
7592 | ||
7593 | static int | |
7594 | inherit_task_group(struct perf_event *event, struct task_struct *parent, | |
7595 | struct perf_event_context *parent_ctx, | |
8dc85d54 | 7596 | struct task_struct *child, int ctxn, |
889ff015 FW |
7597 | int *inherited_all) |
7598 | { | |
7599 | int ret; | |
8dc85d54 | 7600 | struct perf_event_context *child_ctx; |
889ff015 FW |
7601 | |
7602 | if (!event->attr.inherit) { | |
7603 | *inherited_all = 0; | |
7604 | return 0; | |
bbbee908 PZ |
7605 | } |
7606 | ||
fe4b04fa | 7607 | child_ctx = child->perf_event_ctxp[ctxn]; |
889ff015 FW |
7608 | if (!child_ctx) { |
7609 | /* | |
7610 | * This is executed from the parent task context, so | |
7611 | * inherit events that have been marked for cloning. | |
7612 | * First allocate and initialize a context for the | |
7613 | * child. | |
7614 | */ | |
bbbee908 | 7615 | |
734df5ab | 7616 | child_ctx = alloc_perf_context(parent_ctx->pmu, child); |
889ff015 FW |
7617 | if (!child_ctx) |
7618 | return -ENOMEM; | |
bbbee908 | 7619 | |
8dc85d54 | 7620 | child->perf_event_ctxp[ctxn] = child_ctx; |
889ff015 FW |
7621 | } |
7622 | ||
7623 | ret = inherit_group(event, parent, parent_ctx, | |
7624 | child, child_ctx); | |
7625 | ||
7626 | if (ret) | |
7627 | *inherited_all = 0; | |
7628 | ||
7629 | return ret; | |
bbbee908 PZ |
7630 | } |
7631 | ||
9b51f66d | 7632 | /* |
cdd6c482 | 7633 | * Initialize the perf_event context in task_struct |
9b51f66d | 7634 | */ |
8dc85d54 | 7635 | int perf_event_init_context(struct task_struct *child, int ctxn) |
9b51f66d | 7636 | { |
889ff015 | 7637 | struct perf_event_context *child_ctx, *parent_ctx; |
cdd6c482 IM |
7638 | struct perf_event_context *cloned_ctx; |
7639 | struct perf_event *event; | |
9b51f66d | 7640 | struct task_struct *parent = current; |
564c2b21 | 7641 | int inherited_all = 1; |
dddd3379 | 7642 | unsigned long flags; |
6ab423e0 | 7643 | int ret = 0; |
9b51f66d | 7644 | |
8dc85d54 | 7645 | if (likely(!parent->perf_event_ctxp[ctxn])) |
6ab423e0 PZ |
7646 | return 0; |
7647 | ||
ad3a37de | 7648 | /* |
25346b93 PM |
7649 | * If the parent's context is a clone, pin it so it won't get |
7650 | * swapped under us. | |
ad3a37de | 7651 | */ |
8dc85d54 | 7652 | parent_ctx = perf_pin_task_context(parent, ctxn); |
25346b93 | 7653 | |
ad3a37de PM |
7654 | /* |
7655 | * No need to check if parent_ctx != NULL here; since we saw | |
7656 | * it non-NULL earlier, the only reason for it to become NULL | |
7657 | * is if we exit, and since we're currently in the middle of | |
7658 | * a fork we can't be exiting at the same time. | |
7659 | */ | |
ad3a37de | 7660 | |
9b51f66d IM |
7661 | /* |
7662 | * Lock the parent list. No need to lock the child - not PID | |
7663 | * hashed yet and not running, so nobody can access it. | |
7664 | */ | |
d859e29f | 7665 | mutex_lock(&parent_ctx->mutex); |
9b51f66d IM |
7666 | |
7667 | /* | |
7668 | * We dont have to disable NMIs - we are only looking at | |
7669 | * the list, not manipulating it: | |
7670 | */ | |
889ff015 | 7671 | list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) { |
8dc85d54 PZ |
7672 | ret = inherit_task_group(event, parent, parent_ctx, |
7673 | child, ctxn, &inherited_all); | |
889ff015 FW |
7674 | if (ret) |
7675 | break; | |
7676 | } | |
b93f7978 | 7677 | |
dddd3379 TG |
7678 | /* |
7679 | * We can't hold ctx->lock when iterating the ->flexible_group list due | |
7680 | * to allocations, but we need to prevent rotation because | |
7681 | * rotate_ctx() will change the list from interrupt context. | |
7682 | */ | |
7683 | raw_spin_lock_irqsave(&parent_ctx->lock, flags); | |
7684 | parent_ctx->rotate_disable = 1; | |
7685 | raw_spin_unlock_irqrestore(&parent_ctx->lock, flags); | |
7686 | ||
889ff015 | 7687 | list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) { |
8dc85d54 PZ |
7688 | ret = inherit_task_group(event, parent, parent_ctx, |
7689 | child, ctxn, &inherited_all); | |
889ff015 | 7690 | if (ret) |
9b51f66d | 7691 | break; |
564c2b21 PM |
7692 | } |
7693 | ||
dddd3379 TG |
7694 | raw_spin_lock_irqsave(&parent_ctx->lock, flags); |
7695 | parent_ctx->rotate_disable = 0; | |
dddd3379 | 7696 | |
8dc85d54 | 7697 | child_ctx = child->perf_event_ctxp[ctxn]; |
889ff015 | 7698 | |
05cbaa28 | 7699 | if (child_ctx && inherited_all) { |
564c2b21 PM |
7700 | /* |
7701 | * Mark the child context as a clone of the parent | |
7702 | * context, or of whatever the parent is a clone of. | |
c5ed5145 PZ |
7703 | * |
7704 | * Note that if the parent is a clone, the holding of | |
7705 | * parent_ctx->lock avoids it from being uncloned. | |
564c2b21 | 7706 | */ |
c5ed5145 | 7707 | cloned_ctx = parent_ctx->parent_ctx; |
ad3a37de PM |
7708 | if (cloned_ctx) { |
7709 | child_ctx->parent_ctx = cloned_ctx; | |
25346b93 | 7710 | child_ctx->parent_gen = parent_ctx->parent_gen; |
564c2b21 PM |
7711 | } else { |
7712 | child_ctx->parent_ctx = parent_ctx; | |
7713 | child_ctx->parent_gen = parent_ctx->generation; | |
7714 | } | |
7715 | get_ctx(child_ctx->parent_ctx); | |
9b51f66d IM |
7716 | } |
7717 | ||
c5ed5145 | 7718 | raw_spin_unlock_irqrestore(&parent_ctx->lock, flags); |
d859e29f | 7719 | mutex_unlock(&parent_ctx->mutex); |
6ab423e0 | 7720 | |
25346b93 | 7721 | perf_unpin_context(parent_ctx); |
fe4b04fa | 7722 | put_ctx(parent_ctx); |
ad3a37de | 7723 | |
6ab423e0 | 7724 | return ret; |
9b51f66d IM |
7725 | } |
7726 | ||
8dc85d54 PZ |
7727 | /* |
7728 | * Initialize the perf_event context in task_struct | |
7729 | */ | |
7730 | int perf_event_init_task(struct task_struct *child) | |
7731 | { | |
7732 | int ctxn, ret; | |
7733 | ||
8550d7cb ON |
7734 | memset(child->perf_event_ctxp, 0, sizeof(child->perf_event_ctxp)); |
7735 | mutex_init(&child->perf_event_mutex); | |
7736 | INIT_LIST_HEAD(&child->perf_event_list); | |
7737 | ||
8dc85d54 PZ |
7738 | for_each_task_context_nr(ctxn) { |
7739 | ret = perf_event_init_context(child, ctxn); | |
7740 | if (ret) | |
7741 | return ret; | |
7742 | } | |
7743 | ||
7744 | return 0; | |
7745 | } | |
7746 | ||
220b140b PM |
7747 | static void __init perf_event_init_all_cpus(void) |
7748 | { | |
b28ab83c | 7749 | struct swevent_htable *swhash; |
220b140b | 7750 | int cpu; |
220b140b PM |
7751 | |
7752 | for_each_possible_cpu(cpu) { | |
b28ab83c PZ |
7753 | swhash = &per_cpu(swevent_htable, cpu); |
7754 | mutex_init(&swhash->hlist_mutex); | |
e9d2b064 | 7755 | INIT_LIST_HEAD(&per_cpu(rotation_list, cpu)); |
220b140b PM |
7756 | } |
7757 | } | |
7758 | ||
0db0628d | 7759 | static void perf_event_init_cpu(int cpu) |
0793a61d | 7760 | { |
108b02cf | 7761 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
0793a61d | 7762 | |
b28ab83c | 7763 | mutex_lock(&swhash->hlist_mutex); |
4536e4d1 | 7764 | if (swhash->hlist_refcount > 0) { |
76e1d904 FW |
7765 | struct swevent_hlist *hlist; |
7766 | ||
b28ab83c PZ |
7767 | hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu)); |
7768 | WARN_ON(!hlist); | |
7769 | rcu_assign_pointer(swhash->swevent_hlist, hlist); | |
76e1d904 | 7770 | } |
b28ab83c | 7771 | mutex_unlock(&swhash->hlist_mutex); |
0793a61d TG |
7772 | } |
7773 | ||
c277443c | 7774 | #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC |
e9d2b064 | 7775 | static void perf_pmu_rotate_stop(struct pmu *pmu) |
0793a61d | 7776 | { |
e9d2b064 PZ |
7777 | struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context); |
7778 | ||
7779 | WARN_ON(!irqs_disabled()); | |
7780 | ||
7781 | list_del_init(&cpuctx->rotation_list); | |
7782 | } | |
7783 | ||
108b02cf | 7784 | static void __perf_event_exit_context(void *__info) |
0793a61d | 7785 | { |
108b02cf | 7786 | struct perf_event_context *ctx = __info; |
cdd6c482 | 7787 | struct perf_event *event, *tmp; |
0793a61d | 7788 | |
108b02cf | 7789 | perf_pmu_rotate_stop(ctx->pmu); |
b5ab4cd5 | 7790 | |
889ff015 | 7791 | list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry) |
fe4b04fa | 7792 | __perf_remove_from_context(event); |
889ff015 | 7793 | list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry) |
fe4b04fa | 7794 | __perf_remove_from_context(event); |
0793a61d | 7795 | } |
108b02cf PZ |
7796 | |
7797 | static void perf_event_exit_cpu_context(int cpu) | |
7798 | { | |
7799 | struct perf_event_context *ctx; | |
7800 | struct pmu *pmu; | |
7801 | int idx; | |
7802 | ||
7803 | idx = srcu_read_lock(&pmus_srcu); | |
7804 | list_for_each_entry_rcu(pmu, &pmus, entry) { | |
917bdd1c | 7805 | ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx; |
108b02cf PZ |
7806 | |
7807 | mutex_lock(&ctx->mutex); | |
7808 | smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1); | |
7809 | mutex_unlock(&ctx->mutex); | |
7810 | } | |
7811 | srcu_read_unlock(&pmus_srcu, idx); | |
108b02cf PZ |
7812 | } |
7813 | ||
cdd6c482 | 7814 | static void perf_event_exit_cpu(int cpu) |
0793a61d | 7815 | { |
b28ab83c | 7816 | struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu); |
d859e29f | 7817 | |
b28ab83c PZ |
7818 | mutex_lock(&swhash->hlist_mutex); |
7819 | swevent_hlist_release(swhash); | |
7820 | mutex_unlock(&swhash->hlist_mutex); | |
76e1d904 | 7821 | |
108b02cf | 7822 | perf_event_exit_cpu_context(cpu); |
0793a61d TG |
7823 | } |
7824 | #else | |
cdd6c482 | 7825 | static inline void perf_event_exit_cpu(int cpu) { } |
0793a61d TG |
7826 | #endif |
7827 | ||
c277443c PZ |
7828 | static int |
7829 | perf_reboot(struct notifier_block *notifier, unsigned long val, void *v) | |
7830 | { | |
7831 | int cpu; | |
7832 | ||
7833 | for_each_online_cpu(cpu) | |
7834 | perf_event_exit_cpu(cpu); | |
7835 | ||
7836 | return NOTIFY_OK; | |
7837 | } | |
7838 | ||
7839 | /* | |
7840 | * Run the perf reboot notifier at the very last possible moment so that | |
7841 | * the generic watchdog code runs as long as possible. | |
7842 | */ | |
7843 | static struct notifier_block perf_reboot_notifier = { | |
7844 | .notifier_call = perf_reboot, | |
7845 | .priority = INT_MIN, | |
7846 | }; | |
7847 | ||
0db0628d | 7848 | static int |
0793a61d TG |
7849 | perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu) |
7850 | { | |
7851 | unsigned int cpu = (long)hcpu; | |
7852 | ||
4536e4d1 | 7853 | switch (action & ~CPU_TASKS_FROZEN) { |
0793a61d TG |
7854 | |
7855 | case CPU_UP_PREPARE: | |
5e11637e | 7856 | case CPU_DOWN_FAILED: |
cdd6c482 | 7857 | perf_event_init_cpu(cpu); |
0793a61d TG |
7858 | break; |
7859 | ||
5e11637e | 7860 | case CPU_UP_CANCELED: |
0793a61d | 7861 | case CPU_DOWN_PREPARE: |
cdd6c482 | 7862 | perf_event_exit_cpu(cpu); |
0793a61d | 7863 | break; |
0793a61d TG |
7864 | default: |
7865 | break; | |
7866 | } | |
7867 | ||
7868 | return NOTIFY_OK; | |
7869 | } | |
7870 | ||
cdd6c482 | 7871 | void __init perf_event_init(void) |
0793a61d | 7872 | { |
3c502e7a JW |
7873 | int ret; |
7874 | ||
2e80a82a PZ |
7875 | idr_init(&pmu_idr); |
7876 | ||
220b140b | 7877 | perf_event_init_all_cpus(); |
b0a873eb | 7878 | init_srcu_struct(&pmus_srcu); |
2e80a82a PZ |
7879 | perf_pmu_register(&perf_swevent, "software", PERF_TYPE_SOFTWARE); |
7880 | perf_pmu_register(&perf_cpu_clock, NULL, -1); | |
7881 | perf_pmu_register(&perf_task_clock, NULL, -1); | |
b0a873eb PZ |
7882 | perf_tp_register(); |
7883 | perf_cpu_notifier(perf_cpu_notify); | |
c277443c | 7884 | register_reboot_notifier(&perf_reboot_notifier); |
3c502e7a JW |
7885 | |
7886 | ret = init_hw_breakpoint(); | |
7887 | WARN(ret, "hw_breakpoint initialization failed with: %d", ret); | |
b2029520 GN |
7888 | |
7889 | /* do not patch jump label more than once per second */ | |
7890 | jump_label_rate_limit(&perf_sched_events, HZ); | |
b01c3a00 JO |
7891 | |
7892 | /* | |
7893 | * Build time assertion that we keep the data_head at the intended | |
7894 | * location. IOW, validation we got the __reserved[] size right. | |
7895 | */ | |
7896 | BUILD_BUG_ON((offsetof(struct perf_event_mmap_page, data_head)) | |
7897 | != 1024); | |
0793a61d | 7898 | } |
abe43400 PZ |
7899 | |
7900 | static int __init perf_event_sysfs_init(void) | |
7901 | { | |
7902 | struct pmu *pmu; | |
7903 | int ret; | |
7904 | ||
7905 | mutex_lock(&pmus_lock); | |
7906 | ||
7907 | ret = bus_register(&pmu_bus); | |
7908 | if (ret) | |
7909 | goto unlock; | |
7910 | ||
7911 | list_for_each_entry(pmu, &pmus, entry) { | |
7912 | if (!pmu->name || pmu->type < 0) | |
7913 | continue; | |
7914 | ||
7915 | ret = pmu_dev_alloc(pmu); | |
7916 | WARN(ret, "Failed to register pmu: %s, reason %d\n", pmu->name, ret); | |
7917 | } | |
7918 | pmu_bus_running = 1; | |
7919 | ret = 0; | |
7920 | ||
7921 | unlock: | |
7922 | mutex_unlock(&pmus_lock); | |
7923 | ||
7924 | return ret; | |
7925 | } | |
7926 | device_initcall(perf_event_sysfs_init); | |
e5d1367f SE |
7927 | |
7928 | #ifdef CONFIG_CGROUP_PERF | |
eb95419b TH |
7929 | static struct cgroup_subsys_state * |
7930 | perf_cgroup_css_alloc(struct cgroup_subsys_state *parent_css) | |
e5d1367f SE |
7931 | { |
7932 | struct perf_cgroup *jc; | |
e5d1367f | 7933 | |
1b15d055 | 7934 | jc = kzalloc(sizeof(*jc), GFP_KERNEL); |
e5d1367f SE |
7935 | if (!jc) |
7936 | return ERR_PTR(-ENOMEM); | |
7937 | ||
e5d1367f SE |
7938 | jc->info = alloc_percpu(struct perf_cgroup_info); |
7939 | if (!jc->info) { | |
7940 | kfree(jc); | |
7941 | return ERR_PTR(-ENOMEM); | |
7942 | } | |
7943 | ||
e5d1367f SE |
7944 | return &jc->css; |
7945 | } | |
7946 | ||
eb95419b | 7947 | static void perf_cgroup_css_free(struct cgroup_subsys_state *css) |
e5d1367f | 7948 | { |
eb95419b TH |
7949 | struct perf_cgroup *jc = container_of(css, struct perf_cgroup, css); |
7950 | ||
e5d1367f SE |
7951 | free_percpu(jc->info); |
7952 | kfree(jc); | |
7953 | } | |
7954 | ||
7955 | static int __perf_cgroup_move(void *info) | |
7956 | { | |
7957 | struct task_struct *task = info; | |
7958 | perf_cgroup_switch(task, PERF_CGROUP_SWOUT | PERF_CGROUP_SWIN); | |
7959 | return 0; | |
7960 | } | |
7961 | ||
eb95419b TH |
7962 | static void perf_cgroup_attach(struct cgroup_subsys_state *css, |
7963 | struct cgroup_taskset *tset) | |
e5d1367f | 7964 | { |
bb9d97b6 TH |
7965 | struct task_struct *task; |
7966 | ||
d99c8727 | 7967 | cgroup_taskset_for_each(task, css, tset) |
bb9d97b6 | 7968 | task_function_call(task, __perf_cgroup_move, task); |
e5d1367f SE |
7969 | } |
7970 | ||
eb95419b TH |
7971 | static void perf_cgroup_exit(struct cgroup_subsys_state *css, |
7972 | struct cgroup_subsys_state *old_css, | |
761b3ef5 | 7973 | struct task_struct *task) |
e5d1367f SE |
7974 | { |
7975 | /* | |
7976 | * cgroup_exit() is called in the copy_process() failure path. | |
7977 | * Ignore this case since the task hasn't ran yet, this avoids | |
7978 | * trying to poke a half freed task state from generic code. | |
7979 | */ | |
7980 | if (!(task->flags & PF_EXITING)) | |
7981 | return; | |
7982 | ||
bb9d97b6 | 7983 | task_function_call(task, __perf_cgroup_move, task); |
e5d1367f SE |
7984 | } |
7985 | ||
7986 | struct cgroup_subsys perf_subsys = { | |
e7e7ee2e IM |
7987 | .name = "perf_event", |
7988 | .subsys_id = perf_subsys_id, | |
92fb9748 TH |
7989 | .css_alloc = perf_cgroup_css_alloc, |
7990 | .css_free = perf_cgroup_css_free, | |
e7e7ee2e | 7991 | .exit = perf_cgroup_exit, |
bb9d97b6 | 7992 | .attach = perf_cgroup_attach, |
e5d1367f SE |
7993 | }; |
7994 | #endif /* CONFIG_CGROUP_PERF */ |